1	/*M///////////////////////////////////////////////////////////////////////////////////////
2	//
3	// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
4	//
5	// By downloading, copying, installing or using the software you agree to this license.
6	// If you do not agree to this license, do not download, install,
7	// copy or use the software.
8	//
9	//
10	// License Agreement
11	// For Open Source Computer Vision Library
12	//
13	// Copyright (C) 2013, OpenCV Foundation, all rights reserved.
14	// Third party copyrights are property of their respective owners.
15	//
16	// Redistribution and use in source and binary forms, with or without modification,
17	// are permitted provided that the following conditions are met:
18	//
19	// * Redistribution's of source code must retain the above copyright notice,
20	// this list of conditions and the following disclaimer.
21	//
22	// * Redistribution's in binary form must reproduce the above copyright notice,
23	// this list of conditions and the following disclaimer in the documentation
24	// and/or other materials provided with the distribution.
25	//
26	// * The name of the copyright holders may not be used to endorse or promote products
27	// derived from this software without specific prior written permission.
28	//
29	// This software is provided by the copyright holders and contributors "as is" and
30	// any express or implied warranties, including, but not limited to, the implied
31	// warranties of merchantability and fitness for a particular purpose are disclaimed.
32	// In no event shall the Intel Corporation or contributors be liable for any direct,
33	// indirect, incidental, special, exemplary, or consequential damages
34	// (including, but not limited to, procurement of substitute goods or services;
35	// loss of use, data, or profits; or business interruption) however caused
36	// and on any theory of liability, whether in contract, strict liability,
37	// or tort (including negligence or otherwise) arising in any way out of
38	// the use of this software, even if advised of the possibility of such damage.
39	//
40	//M*/
41
42	#ifndef OPENCV_DNN_DNN_HPP
43	#define OPENCV_DNN_DNN_HPP
44
45	#include <vector>
46	#include <opencv2/core.hpp>
47	#include "opencv2/core/async.hpp"
48
49	#include "../dnn/version.hpp"
50
51	#include <opencv2/dnn/dict.hpp>
52
53	namespace cv {
54	namespace dnn {
55
56	namespace accessor {
57	class DnnNetAccessor; // forward declaration
58	}
59
60	CV__DNN_INLINE_NS_BEGIN
61	//! @addtogroup dnn
62	//! @{
63
64	typedef std::vector<int> MatShape;
65
66	/**
67	* @brief Enum of computation backends supported by layers.
68	* @see Net::setPreferableBackend
69	*/
70	enum Backend
71	{
72	//! DNN_BACKEND_DEFAULT equals to OPENCV_DNN_BACKEND_DEFAULT, which can be defined using CMake or a configuration parameter
73	DNN_BACKEND_DEFAULT = 0,
74	DNN_BACKEND_HALIDE,
75	DNN_BACKEND_INFERENCE_ENGINE, //!< Intel OpenVINO computational backend
76	//!< @note Tutorial how to build OpenCV with OpenVINO: @ref tutorial_dnn_openvino
77	DNN_BACKEND_OPENCV,
78	DNN_BACKEND_VKCOM,
79	DNN_BACKEND_CUDA,
80	DNN_BACKEND_WEBNN,
81	DNN_BACKEND_TIMVX,
82	DNN_BACKEND_CANN,
83	#if defined(__OPENCV_BUILD) || defined(BUILD_PLUGIN)
84	#if !defined(OPENCV_BINDING_PARSER)
85	DNN_BACKEND_INFERENCE_ENGINE_NGRAPH = 1000000, // internal - use DNN_BACKEND_INFERENCE_ENGINE + setInferenceEngineBackendType()
86	DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019, // internal - use DNN_BACKEND_INFERENCE_ENGINE + setInferenceEngineBackendType()
87	#endif
88	#endif
89	};
90
91	/**
92	* @brief Enum of target devices for computations.
93	* @see Net::setPreferableTarget
94	*/
95	enum Target
96	{
97	DNN_TARGET_CPU = 0,
98	DNN_TARGET_OPENCL,
99	DNN_TARGET_OPENCL_FP16,
100	DNN_TARGET_MYRIAD,
101	DNN_TARGET_VULKAN,
102	DNN_TARGET_FPGA, //!< FPGA device with CPU fallbacks using Inference Engine's Heterogeneous plugin.
103	DNN_TARGET_CUDA,
104	DNN_TARGET_CUDA_FP16,
105	DNN_TARGET_HDDL,
106	DNN_TARGET_NPU,
107	DNN_TARGET_CPU_FP16, // Only the ARM platform is supported. Low precision computing, accelerate model inference.
108	};
109
110	/**
111	* @brief Enum of data layout for model inference.
112	* @see Image2BlobParams
113	*/
114	enum DataLayout
115	{
116	DNN_LAYOUT_UNKNOWN = 0,
117	DNN_LAYOUT_ND = 1, //!< OpenCV data layout for 2D data.
118	DNN_LAYOUT_NCHW = 2, //!< OpenCV data layout for 4D data.
119	DNN_LAYOUT_NCDHW = 3, //!< OpenCV data layout for 5D data.
120	DNN_LAYOUT_NHWC = 4, //!< Tensorflow-like data layout for 4D data.
121	DNN_LAYOUT_NDHWC = 5, //!< Tensorflow-like data layout for 5D data.
122	DNN_LAYOUT_PLANAR = 6, //!< Tensorflow-like data layout, it should only be used at tf or tflite model parsing.
123	};
124
125	CV_EXPORTS std::vector< std::pair<Backend, Target> > getAvailableBackends();
126	CV_EXPORTS_W std::vector<Target> getAvailableTargets(dnn::Backend be);
127
128	/**
129	* @brief Enables detailed logging of the DNN model loading with CV DNN API.
130	* @param[in] isDiagnosticsMode Indicates whether diagnostic mode should be set.
131	*
132	* Diagnostic mode provides detailed logging of the model loading stage to explore
133	* potential problems (ex.: not implemented layer type).
134	*
135	* @note In diagnostic mode series of assertions will be skipped, it can lead to the
136	* expected application crash.
137	*/
138	CV_EXPORTS void enableModelDiagnostics(bool isDiagnosticsMode);
139
140	/** @brief This class provides all data needed to initialize layer.
141	*
142	* It includes dictionary with scalar params (which can be read by using Dict interface),
143	* blob params #blobs and optional meta information: #name and #type of layer instance.
144	*/
145	class CV_EXPORTS LayerParams : public Dict
146	{
147	public:
148	//TODO: Add ability to name blob params
149	std::vector<Mat> blobs; //!< List of learned parameters stored as blobs.
150
151	String name; //!< Name of the layer instance (optional, can be used internal purposes).
152	String type; //!< Type name which was used for creating layer by layer factory (optional).
153	};
154
155	/**
156	* @brief Derivatives of this class encapsulates functions of certain backends.
157	*/
158	class BackendNode
159	{
160	public:
161	explicit BackendNode(int backendId);
162
163	virtual ~BackendNode(); //!< Virtual destructor to make polymorphism.
164
165	int backendId; //!< Backend identifier.
166	};
167
168	/**
169	* @brief Derivatives of this class wraps cv::Mat for different backends and targets.
170	*/
171	class BackendWrapper
172	{
173	public:
174	BackendWrapper(int backendId, int targetId);
175
176	/**
177	* @brief Wrap cv::Mat for specific backend and target.
178	* @param[in] targetId Target identifier.
179	* @param[in] m cv::Mat for wrapping.
180	*
181	* Make CPU->GPU data transfer if it's require for the target.
182	*/
183	BackendWrapper(int targetId, const cv::Mat& m);
184
185	/**
186	* @brief Make wrapper for reused cv::Mat.
187	* @param[in] base Wrapper of cv::Mat that will be reused.
188	* @param[in] shape Specific shape.
189	*
190	* Initialize wrapper from another one. It'll wrap the same host CPU
191	* memory and mustn't allocate memory on device(i.e. GPU). It might
192	* has different shape. Use in case of CPU memory reusing for reuse
193	* associated memory on device too.
194	*/
195	BackendWrapper(const Ptr<BackendWrapper>& base, const MatShape& shape);
196
197	virtual ~BackendWrapper(); //!< Virtual destructor to make polymorphism.
198
199	/**
200	* @brief Transfer data to CPU host memory.
201	*/
202	virtual void copyToHost() = 0;
203
204	/**
205	* @brief Indicate that an actual data is on CPU.
206	*/
207	virtual void setHostDirty() = 0;
208
209	int backendId; //!< Backend identifier.
210	int targetId; //!< Target identifier.
211	};
212
213	class CV_EXPORTS ActivationLayer;
214
215	/** @brief This interface class allows to build new Layers - are building blocks of networks.
216	*
217	* Each class, derived from Layer, must implement allocate() methods to declare own outputs and forward() to compute outputs.
218	* Also before using the new layer into networks you must register your layer by using one of @ref dnnLayerFactory "LayerFactory" macros.
219	*/
220	class CV_EXPORTS_W Layer : public Algorithm
221	{
222	public:
223
224	//! List of learned parameters must be stored here to allow read them by using Net::getParam().
225	CV_PROP_RW std::vector<Mat> blobs;
226
227	/** @brief Computes and sets internal parameters according to inputs, outputs and blobs.
228	* @deprecated Use Layer::finalize(InputArrayOfArrays, OutputArrayOfArrays) instead
229	* @param[in] input vector of already allocated input blobs
230	* @param[out] output vector of already allocated output blobs
231	*
232	* If this method is called after network has allocated all memory for input and output blobs
233	* and before inferencing.
234	*/
235	CV_DEPRECATED_EXTERNAL
236	virtual void finalize(const std::vector<Mat*> &input, std::vector<Mat> &output);
237
238	/** @brief Computes and sets internal parameters according to inputs, outputs and blobs.
239	* @param[in] inputs vector of already allocated input blobs
240	* @param[out] outputs vector of already allocated output blobs
241	*
242	* If this method is called after network has allocated all memory for input and output blobs
243	* and before inferencing.
244	*/
245	CV_WRAP virtual void finalize(InputArrayOfArrays inputs, OutputArrayOfArrays outputs);
246
247	/** @brief Given the @p input blobs, computes the output @p blobs.
248	* @deprecated Use Layer::forward(InputArrayOfArrays, OutputArrayOfArrays, OutputArrayOfArrays) instead
249	* @param[in] input the input blobs.
250	* @param[out] output allocated output blobs, which will store results of the computation.
251	* @param[out] internals allocated internal blobs
252	*/
253	CV_DEPRECATED_EXTERNAL
254	virtual void forward(std::vector<Mat*> &input, std::vector<Mat> &output, std::vector<Mat> &internals);
255
256	/** @brief Given the @p input blobs, computes the output @p blobs.
257	* @param[in] inputs the input blobs.
258	* @param[out] outputs allocated output blobs, which will store results of the computation.
259	* @param[out] internals allocated internal blobs
260	*/
261	virtual void forward(InputArrayOfArrays inputs, OutputArrayOfArrays outputs, OutputArrayOfArrays internals);
262
263	/** @brief Tries to quantize the given layer and compute the quantization parameters required for fixed point implementation.
264	* @param[in] scales input and output scales.
265	* @param[in] zeropoints input and output zeropoints.
266	* @param[out] params Quantized parameters required for fixed point implementation of that layer.
267	* @returns True if layer can be quantized.
268	*/
269	virtual bool tryQuantize(const std::vector<std::vector<float> > &scales,
270	const std::vector<std::vector<int> > &zeropoints, LayerParams& params);
271
272	/** @brief Given the @p input blobs, computes the output @p blobs.
273	* @param[in] inputs the input blobs.
274	* @param[out] outputs allocated output blobs, which will store results of the computation.
275	* @param[out] internals allocated internal blobs
276	*/
277	void forward_fallback(InputArrayOfArrays inputs, OutputArrayOfArrays outputs, OutputArrayOfArrays internals);
278
279	/** @brief
280	* @overload
281	* @deprecated Use Layer::finalize(InputArrayOfArrays, OutputArrayOfArrays) instead
282	*/
283	CV_DEPRECATED_EXTERNAL
284	void finalize(const std::vector<Mat> &inputs, CV_OUT std::vector<Mat> &outputs);
285
286	/** @brief
287	* @overload
288	* @deprecated Use Layer::finalize(InputArrayOfArrays, OutputArrayOfArrays) instead
289	*/
290	CV_DEPRECATED std::vector<Mat> finalize(const std::vector<Mat> &inputs);
291
292	/** @brief Allocates layer and computes output.
293	* @deprecated This method will be removed in the future release.
294	*/
295	CV_DEPRECATED CV_WRAP void run(const std::vector<Mat> &inputs, CV_OUT std::vector<Mat> &outputs,
296	CV_IN_OUT std::vector<Mat> &internals);
297
298	/** @brief Returns index of input blob into the input array.
299	* @param inputName label of input blob
300	*
301	* Each layer input and output can be labeled to easily identify them using "%<layer_name%>[.output_name]" notation.
302	* This method maps label of input blob to its index into input vector.
303	*/
304	virtual int inputNameToIndex(String inputName); // FIXIT const
305	/** @brief Returns index of output blob in output array.
306	* @see inputNameToIndex()
307	*/
308	CV_WRAP virtual int outputNameToIndex(const String& outputName); // FIXIT const
309
310	/**
311	* @brief Ask layer if it support specific backend for doing computations.
312	* @param[in] backendId computation backend identifier.
313	* @see Backend
314	*/
315	virtual bool supportBackend(int backendId); // FIXIT const
316
317	/**
318	* @brief Returns Halide backend node.
319	* @param[in] inputs Input Halide buffers.
320	* @see BackendNode, BackendWrapper
321	*
322	* Input buffers should be exactly the same that will be used in forward invocations.
323	* Despite we can use Halide::ImageParam based on input shape only,
324	* it helps prevent some memory management issues (if something wrong,
325	* Halide tests will be failed).
326	*/
327	virtual Ptr<BackendNode> initHalide(const std::vector<Ptr<BackendWrapper> > &inputs);
328
329	virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> > &inputs, const std::vector<Ptr<BackendNode> >& nodes);
330
331	virtual Ptr<BackendNode> initVkCom(const std::vector<Ptr<BackendWrapper> > &inputs, std::vector<Ptr<BackendWrapper> > &outputs);
332
333	virtual Ptr<BackendNode> initWebnn(const std::vector<Ptr<BackendWrapper> > &inputs, const std::vector<Ptr<BackendNode> >& nodes);
334
335	/**
336	* @brief Returns a CUDA backend node
337	*
338	* @param context void pointer to CSLContext object
339	* @param inputs layer inputs
340	* @param outputs layer outputs
341	*/
342	virtual Ptr<BackendNode> initCUDA(
343	void *context,
344	const std::vector<Ptr<BackendWrapper>>& inputs,
345	const std::vector<Ptr<BackendWrapper>>& outputs
346	);
347
348	/**
349	* @brief Returns a TimVX backend node
350	*
351	* @param timVxInfo void pointer to CSLContext object
352	* @param inputsWrapper layer inputs
353	* @param outputsWrapper layer outputs
354	* @param isLast if the node is the last one of the TimVX Graph.
355	*/
356	virtual Ptr<BackendNode> initTimVX(void* timVxInfo,
357	const std::vector<Ptr<BackendWrapper> > &inputsWrapper,
358	const std::vector<Ptr<BackendWrapper> > &outputsWrapper,
359	bool isLast);
360
361	/**
362	* @brief Returns a CANN backend node
363	*
364	* @param inputs input tensors of CANN operator
365	* @param outputs output tensors of CANN operator
366	* @param nodes nodes of input tensors
367	*/
368	virtual Ptr<BackendNode> initCann(const std::vector<Ptr<BackendWrapper> > &inputs,
369	const std::vector<Ptr<BackendWrapper> > &outputs,
370	const std::vector<Ptr<BackendNode> >& nodes);
371
372	/**
373	* @brief Automatic Halide scheduling based on layer hyper-parameters.
374	* @param[in] node Backend node with Halide functions.
375	* @param[in] inputs Blobs that will be used in forward invocations.
376	* @param[in] outputs Blobs that will be used in forward invocations.
377	* @param[in] targetId Target identifier
378	* @see BackendNode, Target
379	*
380	* Layer don't use own Halide::Func members because we can have applied
381	* layers fusing. In this way the fused function should be scheduled.
382	*/
383	virtual void applyHalideScheduler(Ptr<BackendNode>& node,
384	const std::vector<Mat*> &inputs,
385	const std::vector<Mat> &outputs,
386	int targetId) const;
387
388	/**
389	* @brief Implement layers fusing.
390	* @param[in] node Backend node of bottom layer.
391	* @see BackendNode
392	*
393	* Actual for graph-based backends. If layer attached successfully,
394	* returns non-empty cv::Ptr to node of the same backend.
395	* Fuse only over the last function.
396	*/
397	virtual Ptr<BackendNode> tryAttach(const Ptr<BackendNode>& node);
398
399	/**
400	* @brief Tries to attach to the layer the subsequent activation layer, i.e. do the layer fusion in a partial case.
401	* @param[in] layer The subsequent activation layer.
402	*
403	* Returns true if the activation layer has been attached successfully.
404	*/
405	virtual bool setActivation(const Ptr<ActivationLayer>& layer);
406
407	/**
408	* @brief Try to fuse current layer with a next one
409	* @param[in] top Next layer to be fused.
410	* @returns True if fusion was performed.
411	*/
412	virtual bool tryFuse(Ptr<Layer>& top);
413
414	/**
415	* @brief Returns parameters of layers with channel-wise multiplication and addition.
416	* @param[out] scale Channel-wise multipliers. Total number of values should
417	* be equal to number of channels.
418	* @param[out] shift Channel-wise offsets. Total number of values should
419	* be equal to number of channels.
420	*
421	* Some layers can fuse their transformations with further layers.
422	* In example, convolution + batch normalization. This way base layer
423	* use weights from layer after it. Fused layer is skipped.
424	* By default, @p scale and @p shift are empty that means layer has no
425	* element-wise multiplications or additions.
426	*/
427	virtual void getScaleShift(Mat& scale, Mat& shift) const;
428
429	/**
430	* @brief Returns scale and zeropoint of layers
431	* @param[out] scale Output scale
432	* @param[out] zeropoint Output zeropoint
433	*
434	* By default, @p scale is 1 and @p zeropoint is 0.
435	*/
436	virtual void getScaleZeropoint(float& scale, int& zeropoint) const;
437
438
439	/**
440	* @brief "Detaches" all the layers, attached to particular layer.
441	*/
442	virtual void unsetAttached();
443
444	virtual bool getMemoryShapes(const std::vector<MatShape> &inputs,
445	const int requiredOutputs,
446	std::vector<MatShape> &outputs,
447	std::vector<MatShape> &internals) const;
448
449	virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
450	const std::vector<MatShape> &outputs) const {CV_UNUSED(inputs); CV_UNUSED(outputs); return 0;}
451
452	virtual bool updateMemoryShapes(const std::vector<MatShape> &inputs);
453
454	CV_PROP String name; //!< Name of the layer instance, can be used for logging or other internal purposes.
455	CV_PROP String type; //!< Type name which was used for creating layer by layer factory.
456	CV_PROP int preferableTarget; //!< prefer target for layer forwarding
457
458	Layer();
459	explicit Layer(const LayerParams &params); //!< Initializes only #name, #type and #blobs fields.
460	void setParamsFrom(const LayerParams &params); //!< Initializes only #name, #type and #blobs fields.
461	virtual ~Layer();
462	};
463
464	/** @brief This class allows to create and manipulate comprehensive artificial neural networks.
465	*
466	* Neural network is presented as directed acyclic graph (DAG), where vertices are Layer instances,
467	* and edges specify relationships between layers inputs and outputs.
468	*
469	* Each network layer has unique integer id and unique string name inside its network.
470	* LayerId can store either layer name or layer id.
471	*
472	* This class supports reference counting of its instances, i. e. copies point to the same instance.
473	*/
474	class CV_EXPORTS_W_SIMPLE Net
475	{
476	public:
477
478	CV_WRAP Net(); //!< Default constructor.
479	CV_WRAP ~Net(); //!< Destructor frees the net only if there aren't references to the net anymore.
480
481	/** @brief Create a network from Intel's Model Optimizer intermediate representation (IR).
482	* @param[in] xml XML configuration file with network's topology.
483	* @param[in] bin Binary file with trained weights.
484	* Networks imported from Intel's Model Optimizer are launched in Intel's Inference Engine
485	* backend.
486	*/
487	CV_WRAP static Net readFromModelOptimizer(const String& xml, const String& bin);
488
489	/** @brief Create a network from Intel's Model Optimizer in-memory buffers with intermediate representation (IR).
490	* @param[in] bufferModelConfig buffer with model's configuration.
491	* @param[in] bufferWeights buffer with model's trained weights.
492	* @returns Net object.
493	*/
494	CV_WRAP static
495	Net readFromModelOptimizer(const std::vector<uchar>& bufferModelConfig, const std::vector<uchar>& bufferWeights);
496
497	/** @brief Create a network from Intel's Model Optimizer in-memory buffers with intermediate representation (IR).
498	* @param[in] bufferModelConfigPtr buffer pointer of model's configuration.
499	* @param[in] bufferModelConfigSize buffer size of model's configuration.
500	* @param[in] bufferWeightsPtr buffer pointer of model's trained weights.
501	* @param[in] bufferWeightsSize buffer size of model's trained weights.
502	* @returns Net object.
503	*/
504	static
505	Net readFromModelOptimizer(const uchar* bufferModelConfigPtr, size_t bufferModelConfigSize,
506	const uchar* bufferWeightsPtr, size_t bufferWeightsSize);
507
508	/** Returns true if there are no layers in the network. */
509	CV_WRAP bool empty() const;
510
511	/** @brief Dump net to String
512	* @returns String with structure, hyperparameters, backend, target and fusion
513	* Call method after setInput(). To see correct backend, target and fusion run after forward().
514	*/
515	CV_WRAP String dump();
516	/** @brief Dump net structure, hyperparameters, backend, target and fusion to dot file
517	* @param path path to output file with .dot extension
518	* @see dump()
519	*/
520	CV_WRAP void dumpToFile(const String& path);
521	/** @brief Adds new layer to the net.
522	* @param name unique name of the adding layer.
523	* @param type typename of the adding layer (type must be registered in LayerRegister).
524	* @param dtype datatype of output blobs.
525	* @param params parameters which will be used to initialize the creating layer.
526	* @returns unique identifier of created layer, or -1 if a failure will happen.
527	*/
528	int addLayer(const String &name, const String &type, const int &dtype, LayerParams &params);
529
530	/** @overload Datatype of output blobs set to default CV_32F */
531	int addLayer(const String &name, const String &type, LayerParams &params);
532
533	/** @brief Adds new layer and connects its first input to the first output of previously added layer.
534	* @see addLayer()
535	*/
536	int addLayerToPrev(const String &name, const String &type, const int &dtype, LayerParams &params);
537
538	/** @overload */
539	int addLayerToPrev(const String &name, const String &type, LayerParams &params);
540
541	/** @brief Converts string name of the layer to the integer identifier.
542	* @returns id of the layer, or -1 if the layer wasn't found.
543	*/
544	CV_WRAP int getLayerId(const String &layer) const;
545
546	CV_WRAP std::vector<String> getLayerNames() const;
547
548	/** @brief Container for strings and integers.
549	*
550	* @deprecated Use getLayerId() with int result.
551	*/
552	typedef DictValue LayerId;
553
554	/** @brief Returns pointer to layer with specified id or name which the network use. */
555	CV_WRAP Ptr<Layer> getLayer(int layerId) const;
556	/** @overload
557	* @deprecated Use int getLayerId(const String &layer)
558	*/
559	CV_WRAP inline Ptr<Layer> getLayer(const String& layerName) const { return getLayer(layerId: getLayerId(layer: layerName)); }
560	/** @overload
561	* @deprecated to be removed
562	*/
563	CV_WRAP Ptr<Layer> getLayer(const LayerId& layerId) const;
564
565	/** @brief Returns pointers to input layers of specific layer. */
566	std::vector<Ptr<Layer> > getLayerInputs(int layerId) const; // FIXIT: CV_WRAP
567
568	/** @brief Connects output of the first layer to input of the second layer.
569	* @param outPin descriptor of the first layer output.
570	* @param inpPin descriptor of the second layer input.
571	*
572	* Descriptors have the following template <DFN>&lt;layer_name&gt;[.input_number]</DFN>:
573	* - the first part of the template <DFN>layer_name</DFN> is string name of the added layer.
574	* If this part is empty then the network input pseudo layer will be used;
575	* - the second optional part of the template <DFN>input_number</DFN>
576	* is either number of the layer input, either label one.
577	* If this part is omitted then the first layer input will be used.
578	*
579	* @see setNetInputs(), Layer::inputNameToIndex(), Layer::outputNameToIndex()
580	*/
581	CV_WRAP void connect(String outPin, String inpPin);
582
583	/** @brief Connects #@p outNum output of the first layer to #@p inNum input of the second layer.
584	* @param outLayerId identifier of the first layer
585	* @param outNum number of the first layer output
586	* @param inpLayerId identifier of the second layer
587	* @param inpNum number of the second layer input
588	*/
589	void connect(int outLayerId, int outNum, int inpLayerId, int inpNum);
590
591	/** @brief Registers network output with name
592	*
593	* Function may create additional 'Identity' layer.
594	*
595	* @param outputName identifier of the output
596	* @param layerId identifier of the second layer
597	* @param outputPort number of the second layer input
598	*
599	* @returns index of bound layer (the same as layerId or newly created)
600	*/
601	int registerOutput(const std::string& outputName, int layerId, int outputPort);
602
603	/** @brief Sets outputs names of the network input pseudo layer.
604	*
605	* Each net always has special own the network input pseudo layer with id=0.
606	* This layer stores the user blobs only and don't make any computations.
607	* In fact, this layer provides the only way to pass user data into the network.
608	* As any other layer, this layer can label its outputs and this function provides an easy way to do this.
609	*/
610	CV_WRAP void setInputsNames(const std::vector<String> &inputBlobNames);
611
612	/** @brief Specify shape of network input.
613	*/
614	CV_WRAP void setInputShape(const String &inputName, const MatShape& shape);
615
616	/** @brief Runs forward pass to compute output of layer with name @p outputName.
617	* @param outputName name for layer which output is needed to get
618	* @return blob for first output of specified layer.
619	* @details By default runs forward pass for the whole network.
620	*/
621	CV_WRAP Mat forward(const String& outputName = String());
622
623	/** @brief Runs forward pass to compute output of layer with name @p outputName.
624	* @param outputName name for layer which output is needed to get
625	* @details By default runs forward pass for the whole network.
626	*
627	* This is an asynchronous version of forward(const String&).
628	* dnn::DNN_BACKEND_INFERENCE_ENGINE backend is required.
629	*/
630	CV_WRAP AsyncArray forwardAsync(const String& outputName = String());
631
632	/** @brief Runs forward pass to compute output of layer with name @p outputName.
633	* @param outputBlobs contains all output blobs for specified layer.
634	* @param outputName name for layer which output is needed to get
635	* @details If @p outputName is empty, runs forward pass for the whole network.
636	*/
637	CV_WRAP void forward(OutputArrayOfArrays outputBlobs, const String& outputName = String());
638
639	/** @brief Runs forward pass to compute outputs of layers listed in @p outBlobNames.
640	* @param outputBlobs contains blobs for first outputs of specified layers.
641	* @param outBlobNames names for layers which outputs are needed to get
642	*/
643	CV_WRAP void forward(OutputArrayOfArrays outputBlobs,
644	const std::vector<String>& outBlobNames);
645
646	/** @brief Runs forward pass to compute outputs of layers listed in @p outBlobNames.
647	* @param outputBlobs contains all output blobs for each layer specified in @p outBlobNames.
648	* @param outBlobNames names for layers which outputs are needed to get
649	*/
650	CV_WRAP_AS(forwardAndRetrieve) void forward(CV_OUT std::vector<std::vector<Mat> >& outputBlobs,
651	const std::vector<String>& outBlobNames);
652
653	/** @brief Returns a quantized Net from a floating-point Net.
654	* @param calibData Calibration data to compute the quantization parameters.
655	* @param inputsDtype Datatype of quantized net's inputs. Can be CV_32F or CV_8S.
656	* @param outputsDtype Datatype of quantized net's outputs. Can be CV_32F or CV_8S.
657	* @param perChannel Quantization granularity of quantized Net. The default is true, that means quantize model
658	* in per-channel way (channel-wise). Set it false to quantize model in per-tensor way (or tensor-wise).
659	*/
660	CV_WRAP Net quantize(InputArrayOfArrays calibData, int inputsDtype, int outputsDtype, bool perChannel=true);
661
662	/** @brief Returns input scale and zeropoint for a quantized Net.
663	* @param scales output parameter for returning input scales.
664	* @param zeropoints output parameter for returning input zeropoints.
665	*/
666	CV_WRAP void getInputDetails(CV_OUT std::vector<float>& scales, CV_OUT std::vector<int>& zeropoints) const;
667
668	/** @brief Returns output scale and zeropoint for a quantized Net.
669	* @param scales output parameter for returning output scales.
670	* @param zeropoints output parameter for returning output zeropoints.
671	*/
672	CV_WRAP void getOutputDetails(CV_OUT std::vector<float>& scales, CV_OUT std::vector<int>& zeropoints) const;
673
674	/**
675	* @brief Compile Halide layers.
676	* @param[in] scheduler Path to YAML file with scheduling directives.
677	* @see setPreferableBackend
678	*
679	* Schedule layers that support Halide backend. Then compile them for
680	* specific target. For layers that not represented in scheduling file
681	* or if no manual scheduling used at all, automatic scheduling will be applied.
682	*/
683	CV_WRAP void setHalideScheduler(const String& scheduler);
684
685	/**
686	* @brief Ask network to use specific computation backend where it supported.
687	* @param[in] backendId backend identifier.
688	* @see Backend
689	*/
690	CV_WRAP void setPreferableBackend(int backendId);
691
692	/**
693	* @brief Ask network to make computations on specific target device.
694	* @param[in] targetId target identifier.
695	* @see Target
696	*
697	* List of supported combinations backend / target:
698	* | | DNN_BACKEND_OPENCV | DNN_BACKEND_INFERENCE_ENGINE | DNN_BACKEND_HALIDE | DNN_BACKEND_CUDA |
699	* |------------------------|--------------------|------------------------------|--------------------|-------------------|
700	* | DNN_TARGET_CPU | + | + | + | |
701	* | DNN_TARGET_OPENCL | + | + | + | |
702	* | DNN_TARGET_OPENCL_FP16 | + | + | | |
703	* | DNN_TARGET_MYRIAD | | + | | |
704	* | DNN_TARGET_FPGA | | + | | |
705	* | DNN_TARGET_CUDA | | | | + |
706	* | DNN_TARGET_CUDA_FP16 | | | | + |
707	* | DNN_TARGET_HDDL | | + | | |
708	*/
709	CV_WRAP void setPreferableTarget(int targetId);
710
711	/** @brief Sets the new input value for the network
712	* @param blob A new blob. Should have CV_32F or CV_8U depth.
713	* @param name A name of input layer.
714	* @param scalefactor An optional normalization scale.
715	* @param mean An optional mean subtraction values.
716	* @see connect(String, String) to know format of the descriptor.
717	*
718	* If scale or mean values are specified, a final input blob is computed
719	* as:
720	* \f[input(n,c,h,w) = scalefactor \times (blob(n,c,h,w) - mean_c)\f]
721	*/
722	CV_WRAP void setInput(InputArray blob, const String& name = "",
723	double scalefactor = 1.0, const Scalar& mean = Scalar());
724
725	/** @brief Sets the new value for the learned param of the layer.
726	* @param layer name or id of the layer.
727	* @param numParam index of the layer parameter in the Layer::blobs array.
728	* @param blob the new value.
729	* @see Layer::blobs
730	* @note If shape of the new blob differs from the previous shape,
731	* then the following forward pass may fail.
732	*/
733	CV_WRAP void setParam(int layer, int numParam, const Mat &blob);
734	CV_WRAP inline void setParam(const String& layerName, int numParam, const Mat &blob) { return setParam(layer: getLayerId(layer: layerName), numParam, blob); }
735
736	/** @brief Returns parameter blob of the layer.
737	* @param layer name or id of the layer.
738	* @param numParam index of the layer parameter in the Layer::blobs array.
739	* @see Layer::blobs
740	*/
741	CV_WRAP Mat getParam(int layer, int numParam = 0) const;
742	CV_WRAP inline Mat getParam(const String& layerName, int numParam = 0) const { return getParam(layer: getLayerId(layer: layerName), numParam); }
743
744	/** @brief Returns indexes of layers with unconnected outputs.
745	*
746	* FIXIT: Rework API to registerOutput() approach, deprecate this call
747	*/
748	CV_WRAP std::vector<int> getUnconnectedOutLayers() const;
749
750	/** @brief Returns names of layers with unconnected outputs.
751	*
752	* FIXIT: Rework API to registerOutput() approach, deprecate this call
753	*/
754	CV_WRAP std::vector<String> getUnconnectedOutLayersNames() const;
755
756	/** @brief Returns input and output shapes for all layers in loaded model;
757	* preliminary inferencing isn't necessary.
758	* @param netInputShapes shapes for all input blobs in net input layer.
759	* @param layersIds output parameter for layer IDs.
760	* @param inLayersShapes output parameter for input layers shapes;
761	* order is the same as in layersIds
762	* @param outLayersShapes output parameter for output layers shapes;
763	* order is the same as in layersIds
764	*/
765	CV_WRAP void getLayersShapes(const std::vector<MatShape>& netInputShapes,
766	CV_OUT std::vector<int>& layersIds,
767	CV_OUT std::vector<std::vector<MatShape> >& inLayersShapes,
768	CV_OUT std::vector<std::vector<MatShape> >& outLayersShapes) const;
769
770	/** @overload */
771	CV_WRAP void getLayersShapes(const MatShape& netInputShape,
772	CV_OUT std::vector<int>& layersIds,
773	CV_OUT std::vector<std::vector<MatShape> >& inLayersShapes,
774	CV_OUT std::vector<std::vector<MatShape> >& outLayersShapes) const;
775
776	/** @brief Returns input and output shapes for layer with specified
777	* id in loaded model; preliminary inferencing isn't necessary.
778	* @param netInputShape shape input blob in net input layer.
779	* @param layerId id for layer.
780	* @param inLayerShapes output parameter for input layers shapes;
781	* order is the same as in layersIds
782	* @param outLayerShapes output parameter for output layers shapes;
783	* order is the same as in layersIds
784	*/
785	void getLayerShapes(const MatShape& netInputShape,
786	const int layerId,
787	CV_OUT std::vector<MatShape>& inLayerShapes,
788	CV_OUT std::vector<MatShape>& outLayerShapes) const; // FIXIT: CV_WRAP
789
790	/** @overload */
791	void getLayerShapes(const std::vector<MatShape>& netInputShapes,
792	const int layerId,
793	CV_OUT std::vector<MatShape>& inLayerShapes,
794	CV_OUT std::vector<MatShape>& outLayerShapes) const; // FIXIT: CV_WRAP
795
796	/** @brief Computes FLOP for whole loaded model with specified input shapes.
797	* @param netInputShapes vector of shapes for all net inputs.
798	* @returns computed FLOP.
799	*/
800	CV_WRAP int64 getFLOPS(const std::vector<MatShape>& netInputShapes) const;
801	/** @overload */
802	CV_WRAP int64 getFLOPS(const MatShape& netInputShape) const;
803	/** @overload */
804	CV_WRAP int64 getFLOPS(const int layerId,
805	const std::vector<MatShape>& netInputShapes) const;
806	/** @overload */
807	CV_WRAP int64 getFLOPS(const int layerId,
808	const MatShape& netInputShape) const;
809
810	/** @brief Returns list of types for layer used in model.
811	* @param layersTypes output parameter for returning types.
812	*/
813	CV_WRAP void getLayerTypes(CV_OUT std::vector<String>& layersTypes) const;
814
815	/** @brief Returns count of layers of specified type.
816	* @param layerType type.
817	* @returns count of layers
818	*/
819	CV_WRAP int getLayersCount(const String& layerType) const;
820
821	/** @brief Computes bytes number which are required to store
822	* all weights and intermediate blobs for model.
823	* @param netInputShapes vector of shapes for all net inputs.
824	* @param weights output parameter to store resulting bytes for weights.
825	* @param blobs output parameter to store resulting bytes for intermediate blobs.
826	*/
827	void getMemoryConsumption(const std::vector<MatShape>& netInputShapes,
828	CV_OUT size_t& weights, CV_OUT size_t& blobs) const; // FIXIT: CV_WRAP
829	/** @overload */
830	CV_WRAP void getMemoryConsumption(const MatShape& netInputShape,
831	CV_OUT size_t& weights, CV_OUT size_t& blobs) const;
832	/** @overload */
833	CV_WRAP void getMemoryConsumption(const int layerId,
834	const std::vector<MatShape>& netInputShapes,
835	CV_OUT size_t& weights, CV_OUT size_t& blobs) const;
836	/** @overload */
837	CV_WRAP void getMemoryConsumption(const int layerId,
838	const MatShape& netInputShape,
839	CV_OUT size_t& weights, CV_OUT size_t& blobs) const;
840
841	/** @brief Computes bytes number which are required to store
842	* all weights and intermediate blobs for each layer.
843	* @param netInputShapes vector of shapes for all net inputs.
844	* @param layerIds output vector to save layer IDs.
845	* @param weights output parameter to store resulting bytes for weights.
846	* @param blobs output parameter to store resulting bytes for intermediate blobs.
847	*/
848	void getMemoryConsumption(const std::vector<MatShape>& netInputShapes,
849	CV_OUT std::vector<int>& layerIds,
850	CV_OUT std::vector<size_t>& weights,
851	CV_OUT std::vector<size_t>& blobs) const; // FIXIT: CV_WRAP
852	/** @overload */
853	void getMemoryConsumption(const MatShape& netInputShape,
854	CV_OUT std::vector<int>& layerIds,
855	CV_OUT std::vector<size_t>& weights,
856	CV_OUT std::vector<size_t>& blobs) const; // FIXIT: CV_WRAP
857
858	/** @brief Enables or disables layer fusion in the network.
859	* @param fusion true to enable the fusion, false to disable. The fusion is enabled by default.
860	*/
861	CV_WRAP void enableFusion(bool fusion);
862
863	/** @brief Enables or disables the Winograd compute branch. The Winograd compute branch can speed up
864	* 3x3 Convolution at a small loss of accuracy.
865	* @param useWinograd true to enable the Winograd compute branch. The default is true.
866	*/
867	CV_WRAP void enableWinograd(bool useWinograd);
868
869	/** @brief Returns overall time for inference and timings (in ticks) for layers.
870	*
871	* Indexes in returned vector correspond to layers ids. Some layers can be fused with others,
872	* in this case zero ticks count will be return for that skipped layers. Supported by DNN_BACKEND_OPENCV on DNN_TARGET_CPU only.
873	*
874	* @param[out] timings vector for tick timings for all layers.
875	* @return overall ticks for model inference.
876	*/
877	CV_WRAP int64 getPerfProfile(CV_OUT std::vector<double>& timings);
878
879
880	struct Impl;
881	inline Impl* getImpl() const { return impl.get(); }
882	inline Impl& getImplRef() const { CV_DbgAssert(impl); return *impl.get(); }
883	friend class accessor::DnnNetAccessor;
884	protected:
885	Ptr<Impl> impl;
886	};
887
888	/** @brief Reads a network model stored in <a href="https://pjreddie.com/darknet/">Darknet</a> model files.
889	* @param cfgFile path to the .cfg file with text description of the network architecture.
890	* @param darknetModel path to the .weights file with learned network.
891	* @returns Network object that ready to do forward, throw an exception in failure cases.
892	*/
893	CV_EXPORTS_W Net readNetFromDarknet(const String &cfgFile, const String &darknetModel = String());
894
895	/** @brief Reads a network model stored in <a href="https://pjreddie.com/darknet/">Darknet</a> model files.
896	* @param bufferCfg A buffer contains a content of .cfg file with text description of the network architecture.
897	* @param bufferModel A buffer contains a content of .weights file with learned network.
898	* @returns Net object.
899	*/
900	CV_EXPORTS_W Net readNetFromDarknet(const std::vector<uchar>& bufferCfg,
901	const std::vector<uchar>& bufferModel = std::vector<uchar>());
902
903	/** @brief Reads a network model stored in <a href="https://pjreddie.com/darknet/">Darknet</a> model files.
904	* @param bufferCfg A buffer contains a content of .cfg file with text description of the network architecture.
905	* @param lenCfg Number of bytes to read from bufferCfg
906	* @param bufferModel A buffer contains a content of .weights file with learned network.
907	* @param lenModel Number of bytes to read from bufferModel
908	* @returns Net object.
909	*/
910	CV_EXPORTS Net readNetFromDarknet(const char *bufferCfg, size_t lenCfg,
911	const char *bufferModel = NULL, size_t lenModel = 0);
912
913	/** @brief Reads a network model stored in <a href="http://caffe.berkeleyvision.org">Caffe</a> framework's format.
914	* @param prototxt path to the .prototxt file with text description of the network architecture.
915	* @param caffeModel path to the .caffemodel file with learned network.
916	* @returns Net object.
917	*/
918	CV_EXPORTS_W Net readNetFromCaffe(const String &prototxt, const String &caffeModel = String());
919
920	/** @brief Reads a network model stored in Caffe model in memory.
921	* @param bufferProto buffer containing the content of the .prototxt file
922	* @param bufferModel buffer containing the content of the .caffemodel file
923	* @returns Net object.
924	*/
925	CV_EXPORTS_W Net readNetFromCaffe(const std::vector<uchar>& bufferProto,
926	const std::vector<uchar>& bufferModel = std::vector<uchar>());
927
928	/** @brief Reads a network model stored in Caffe model in memory.
929	* @details This is an overloaded member function, provided for convenience.
930	* It differs from the above function only in what argument(s) it accepts.
931	* @param bufferProto buffer containing the content of the .prototxt file
932	* @param lenProto length of bufferProto
933	* @param bufferModel buffer containing the content of the .caffemodel file
934	* @param lenModel length of bufferModel
935	* @returns Net object.
936	*/
937	CV_EXPORTS Net readNetFromCaffe(const char *bufferProto, size_t lenProto,
938	const char *bufferModel = NULL, size_t lenModel = 0);
939
940	/** @brief Reads a network model stored in <a href="https://www.tensorflow.org/">TensorFlow</a> framework's format.
941	* @param model path to the .pb file with binary protobuf description of the network architecture
942	* @param config path to the .pbtxt file that contains text graph definition in protobuf format.
943	* Resulting Net object is built by text graph using weights from a binary one that
944	* let us make it more flexible.
945	* @returns Net object.
946	*/
947	CV_EXPORTS_W Net readNetFromTensorflow(const String &model, const String &config = String());
948
949	/** @brief Reads a network model stored in <a href="https://www.tensorflow.org/">TensorFlow</a> framework's format.
950	* @param bufferModel buffer containing the content of the pb file
951	* @param bufferConfig buffer containing the content of the pbtxt file
952	* @returns Net object.
953	*/
954	CV_EXPORTS_W Net readNetFromTensorflow(const std::vector<uchar>& bufferModel,
955	const std::vector<uchar>& bufferConfig = std::vector<uchar>());
956
957	/** @brief Reads a network model stored in <a href="https://www.tensorflow.org/">TensorFlow</a> framework's format.
958	* @details This is an overloaded member function, provided for convenience.
959	* It differs from the above function only in what argument(s) it accepts.
960	* @param bufferModel buffer containing the content of the pb file
961	* @param lenModel length of bufferModel
962	* @param bufferConfig buffer containing the content of the pbtxt file
963	* @param lenConfig length of bufferConfig
964	*/
965	CV_EXPORTS Net readNetFromTensorflow(const char *bufferModel, size_t lenModel,
966	const char *bufferConfig = NULL, size_t lenConfig = 0);
967
968	/** @brief Reads a network model stored in <a href="https://www.tensorflow.org/lite">TFLite</a> framework's format.
969	* @param model path to the .tflite file with binary flatbuffers description of the network architecture
970	* @returns Net object.
971	*/
972	CV_EXPORTS_W Net readNetFromTFLite(const String &model);
973
974	/** @brief Reads a network model stored in <a href="https://www.tensorflow.org/lite">TFLite</a> framework's format.
975	* @param bufferModel buffer containing the content of the tflite file
976	* @returns Net object.
977	*/
978	CV_EXPORTS_W Net readNetFromTFLite(const std::vector<uchar>& bufferModel);
979
980	/** @brief Reads a network model stored in <a href="https://www.tensorflow.org/lite">TFLite</a> framework's format.
981	* @details This is an overloaded member function, provided for convenience.
982	* It differs from the above function only in what argument(s) it accepts.
983	* @param bufferModel buffer containing the content of the tflite file
984	* @param lenModel length of bufferModel
985	*/
986	CV_EXPORTS Net readNetFromTFLite(const char *bufferModel, size_t lenModel);
987
988	/**
989	* @brief Reads a network model stored in <a href="http://torch.ch">Torch7</a> framework's format.
990	* @param model path to the file, dumped from Torch by using torch.save() function.
991	* @param isBinary specifies whether the network was serialized in ascii mode or binary.
992	* @param evaluate specifies testing phase of network. If true, it's similar to evaluate() method in Torch.
993	* @returns Net object.
994	*
995	* @note Ascii mode of Torch serializer is more preferable, because binary mode extensively use `long` type of C language,
996	* which has various bit-length on different systems.
997	*
998	* The loading file must contain serialized <a href="https://github.com/torch/nn/blob/master/doc/module.md">nn.Module</a> object
999	* with importing network. Try to eliminate a custom objects from serialazing data to avoid importing errors.
1000	*
1001	* List of supported layers (i.e. object instances derived from Torch nn.Module class):
1002	* - nn.Sequential
1003	* - nn.Parallel
1004	* - nn.Concat
1005	* - nn.Linear
1006	* - nn.SpatialConvolution
1007	* - nn.SpatialMaxPooling, nn.SpatialAveragePooling
1008	* - nn.ReLU, nn.TanH, nn.Sigmoid
1009	* - nn.Reshape
1010	* - nn.SoftMax, nn.LogSoftMax
1011	*
1012	* Also some equivalents of these classes from cunn, cudnn, and fbcunn may be successfully imported.
1013	*/
1014	CV_EXPORTS_W Net readNetFromTorch(const String &model, bool isBinary = true, bool evaluate = true);
1015
1016	/**
1017	* @brief Read deep learning network represented in one of the supported formats.
1018	* @param[in] model Binary file contains trained weights. The following file
1019	* extensions are expected for models from different frameworks:
1020	* * `*.caffemodel` (Caffe, http://caffe.berkeleyvision.org/)
1021	* * `*.pb` (TensorFlow, https://www.tensorflow.org/)
1022	* * `*.t7` | `*.net` (Torch, http://torch.ch/)
1023	* * `*.weights` (Darknet, https://pjreddie.com/darknet/)
1024	* * `*.bin` | `*.onnx` (OpenVINO, https://software.intel.com/openvino-toolkit)
1025	* * `*.onnx` (ONNX, https://onnx.ai/)
1026	* @param[in] config Text file contains network configuration. It could be a
1027	* file with the following extensions:
1028	* * `*.prototxt` (Caffe, http://caffe.berkeleyvision.org/)
1029	* * `*.pbtxt` (TensorFlow, https://www.tensorflow.org/)
1030	* * `*.cfg` (Darknet, https://pjreddie.com/darknet/)
1031	* * `*.xml` (OpenVINO, https://software.intel.com/openvino-toolkit)
1032	* @param[in] framework Explicit framework name tag to determine a format.
1033	* @returns Net object.
1034	*
1035	* This function automatically detects an origin framework of trained model
1036	* and calls an appropriate function such @ref readNetFromCaffe, @ref readNetFromTensorflow,
1037	* @ref readNetFromTorch or @ref readNetFromDarknet. An order of @p model and @p config
1038	* arguments does not matter.
1039	*/
1040	CV_EXPORTS_W Net readNet(const String& model, const String& config = "", const String& framework = "");
1041
1042	/**
1043	* @brief Read deep learning network represented in one of the supported formats.
1044	* @details This is an overloaded member function, provided for convenience.
1045	* It differs from the above function only in what argument(s) it accepts.
1046	* @param[in] framework Name of origin framework.
1047	* @param[in] bufferModel A buffer with a content of binary file with weights
1048	* @param[in] bufferConfig A buffer with a content of text file contains network configuration.
1049	* @returns Net object.
1050	*/
1051	CV_EXPORTS_W Net readNet(const String& framework, const std::vector<uchar>& bufferModel,
1052	const std::vector<uchar>& bufferConfig = std::vector<uchar>());
1053
1054	/** @brief Loads blob which was serialized as torch.Tensor object of Torch7 framework.
1055	* @warning This function has the same limitations as readNetFromTorch().
1056	*/
1057	CV_EXPORTS_W Mat readTorchBlob(const String &filename, bool isBinary = true);
1058
1059	/** @brief Load a network from Intel's Model Optimizer intermediate representation.
1060	* @param[in] xml XML configuration file with network's topology.
1061	* @param[in] bin Binary file with trained weights.
1062	* @returns Net object.
1063	* Networks imported from Intel's Model Optimizer are launched in Intel's Inference Engine
1064	* backend.
1065	*/
1066	CV_EXPORTS_W
1067	Net readNetFromModelOptimizer(const String &xml, const String &bin = "");
1068
1069	/** @brief Load a network from Intel's Model Optimizer intermediate representation.
1070	* @param[in] bufferModelConfig Buffer contains XML configuration with network's topology.
1071	* @param[in] bufferWeights Buffer contains binary data with trained weights.
1072	* @returns Net object.
1073	* Networks imported from Intel's Model Optimizer are launched in Intel's Inference Engine
1074	* backend.
1075	*/
1076	CV_EXPORTS_W
1077	Net readNetFromModelOptimizer(const std::vector<uchar>& bufferModelConfig, const std::vector<uchar>& bufferWeights);
1078
1079	/** @brief Load a network from Intel's Model Optimizer intermediate representation.
1080	* @param[in] bufferModelConfigPtr Pointer to buffer which contains XML configuration with network's topology.
1081	* @param[in] bufferModelConfigSize Binary size of XML configuration data.
1082	* @param[in] bufferWeightsPtr Pointer to buffer which contains binary data with trained weights.
1083	* @param[in] bufferWeightsSize Binary size of trained weights data.
1084	* @returns Net object.
1085	* Networks imported from Intel's Model Optimizer are launched in Intel's Inference Engine
1086	* backend.
1087	*/
1088	CV_EXPORTS
1089	Net readNetFromModelOptimizer(const uchar* bufferModelConfigPtr, size_t bufferModelConfigSize,
1090	const uchar* bufferWeightsPtr, size_t bufferWeightsSize);
1091
1092	/** @brief Reads a network model <a href="https://onnx.ai/">ONNX</a>.
1093	* @param onnxFile path to the .onnx file with text description of the network architecture.
1094	* @returns Network object that ready to do forward, throw an exception in failure cases.
1095	*/
1096	CV_EXPORTS_W Net readNetFromONNX(const String &onnxFile);
1097
1098	/** @brief Reads a network model from <a href="https://onnx.ai/">ONNX</a>
1099	* in-memory buffer.
1100	* @param buffer memory address of the first byte of the buffer.
1101	* @param sizeBuffer size of the buffer.
1102	* @returns Network object that ready to do forward, throw an exception
1103	* in failure cases.
1104	*/
1105	CV_EXPORTS Net readNetFromONNX(const char* buffer, size_t sizeBuffer);
1106
1107	/** @brief Reads a network model from <a href="https://onnx.ai/">ONNX</a>
1108	* in-memory buffer.
1109	* @param buffer in-memory buffer that stores the ONNX model bytes.
1110	* @returns Network object that ready to do forward, throw an exception
1111	* in failure cases.
1112	*/
1113	CV_EXPORTS_W Net readNetFromONNX(const std::vector<uchar>& buffer);
1114
1115	/** @brief Creates blob from .pb file.
1116	* @param path to the .pb file with input tensor.
1117	* @returns Mat.
1118	*/
1119	CV_EXPORTS_W Mat readTensorFromONNX(const String& path);
1120
1121	/** @brief Creates 4-dimensional blob from image. Optionally resizes and crops @p image from center,
1122	* subtract @p mean values, scales values by @p scalefactor, swap Blue and Red channels.
1123	* @param image input image (with 1-, 3- or 4-channels).
1124	* @param scalefactor multiplier for @p images values.
1125	* @param size spatial size for output image
1126	* @param mean scalar with mean values which are subtracted from channels. Values are intended
1127	* to be in (mean-R, mean-G, mean-B) order if @p image has BGR ordering and @p swapRB is true.
1128	* @param swapRB flag which indicates that swap first and last channels
1129	* in 3-channel image is necessary.
1130	* @param crop flag which indicates whether image will be cropped after resize or not
1131	* @param ddepth Depth of output blob. Choose CV_32F or CV_8U.
1132	* @details if @p crop is true, input image is resized so one side after resize is equal to corresponding
1133	* dimension in @p size and another one is equal or larger. Then, crop from the center is performed.
1134	* If @p crop is false, direct resize without cropping and preserving aspect ratio is performed.
1135	* @returns 4-dimensional Mat with NCHW dimensions order.
1136	*
1137	* @note
1138	* The order and usage of `scalefactor` and `mean` are (input - mean) * scalefactor.
1139	*/
1140	CV_EXPORTS_W Mat blobFromImage(InputArray image, double scalefactor=1.0, const Size& size = Size(),
1141	const Scalar& mean = Scalar(), bool swapRB=false, bool crop=false,
1142	int ddepth=CV_32F);
1143
1144	/** @brief Creates 4-dimensional blob from image.
1145	* @details This is an overloaded member function, provided for convenience.
1146	* It differs from the above function only in what argument(s) it accepts.
1147	*/
1148	CV_EXPORTS void blobFromImage(InputArray image, OutputArray blob, double scalefactor=1.0,
1149	const Size& size = Size(), const Scalar& mean = Scalar(),
1150	bool swapRB=false, bool crop=false, int ddepth=CV_32F);
1151
1152
1153	/** @brief Creates 4-dimensional blob from series of images. Optionally resizes and
1154	* crops @p images from center, subtract @p mean values, scales values by @p scalefactor,
1155	* swap Blue and Red channels.
1156	* @param images input images (all with 1-, 3- or 4-channels).
1157	* @param size spatial size for output image
1158	* @param mean scalar with mean values which are subtracted from channels. Values are intended
1159	* to be in (mean-R, mean-G, mean-B) order if @p image has BGR ordering and @p swapRB is true.
1160	* @param scalefactor multiplier for @p images values.
1161	* @param swapRB flag which indicates that swap first and last channels
1162	* in 3-channel image is necessary.
1163	* @param crop flag which indicates whether image will be cropped after resize or not
1164	* @param ddepth Depth of output blob. Choose CV_32F or CV_8U.
1165	* @details if @p crop is true, input image is resized so one side after resize is equal to corresponding
1166	* dimension in @p size and another one is equal or larger. Then, crop from the center is performed.
1167	* If @p crop is false, direct resize without cropping and preserving aspect ratio is performed.
1168	* @returns 4-dimensional Mat with NCHW dimensions order.
1169	*
1170	* @note
1171	* The order and usage of `scalefactor` and `mean` are (input - mean) * scalefactor.
1172	*/
1173	CV_EXPORTS_W Mat blobFromImages(InputArrayOfArrays images, double scalefactor=1.0,
1174	Size size = Size(), const Scalar& mean = Scalar(), bool swapRB=false, bool crop=false,
1175	int ddepth=CV_32F);
1176
1177	/** @brief Creates 4-dimensional blob from series of images.
1178	* @details This is an overloaded member function, provided for convenience.
1179	* It differs from the above function only in what argument(s) it accepts.
1180	*/
1181	CV_EXPORTS void blobFromImages(InputArrayOfArrays images, OutputArray blob,
1182	double scalefactor=1.0, Size size = Size(),
1183	const Scalar& mean = Scalar(), bool swapRB=false, bool crop=false,
1184	int ddepth=CV_32F);
1185
1186	/**
1187	* @brief Enum of image processing mode.
1188	* To facilitate the specialization pre-processing requirements of the dnn model.
1189	* For example, the `letter box` often used in the Yolo series of models.
1190	* @see Image2BlobParams
1191	*/
1192	enum ImagePaddingMode
1193	{
1194	DNN_PMODE_NULL = 0, // !< Default. Resize to required input size without extra processing.
1195	DNN_PMODE_CROP_CENTER = 1, // !< Image will be cropped after resize.
1196	DNN_PMODE_LETTERBOX = 2, // !< Resize image to the desired size while preserving the aspect ratio of original image.
1197	};
1198
1199	/** @brief Processing params of image to blob.
1200	*
1201	* It includes all possible image processing operations and corresponding parameters.
1202	*
1203	* @see blobFromImageWithParams
1204	*
1205	* @note
1206	* The order and usage of `scalefactor` and `mean` are (input - mean) * scalefactor.
1207	* The order and usage of `scalefactor`, `size`, `mean`, `swapRB`, and `ddepth` are consistent
1208	* with the function of @ref blobFromImage.
1209	*/
1210	struct CV_EXPORTS_W_SIMPLE Image2BlobParams
1211	{
1212	CV_WRAP Image2BlobParams();
1213	CV_WRAP Image2BlobParams(const Scalar& scalefactor, const Size& size = Size(), const Scalar& mean = Scalar(),
1214	bool swapRB = false, int ddepth = CV_32F, DataLayout datalayout = DNN_LAYOUT_NCHW,
1215	ImagePaddingMode mode = DNN_PMODE_NULL, Scalar borderValue = 0.0);
1216
1217	CV_PROP_RW Scalar scalefactor; //!< scalefactor multiplier for input image values.
1218	CV_PROP_RW Size size; //!< Spatial size for output image.
1219	CV_PROP_RW Scalar mean; //!< Scalar with mean values which are subtracted from channels.
1220	CV_PROP_RW bool swapRB; //!< Flag which indicates that swap first and last channels
1221	CV_PROP_RW int ddepth; //!< Depth of output blob. Choose CV_32F or CV_8U.
1222	CV_PROP_RW DataLayout datalayout; //!< Order of output dimensions. Choose DNN_LAYOUT_NCHW or DNN_LAYOUT_NHWC.
1223	CV_PROP_RW ImagePaddingMode paddingmode; //!< Image padding mode. @see ImagePaddingMode.
1224	CV_PROP_RW Scalar borderValue; //!< Value used in padding mode for padding.
1225
1226	/** @brief Get rectangle coordinates in original image system from rectangle in blob coordinates.
1227	* @param rBlob rect in blob coordinates.
1228	* @param size original input image size.
1229	* @returns rectangle in original image coordinates.
1230	*/
1231	CV_WRAP Rect blobRectToImageRect(const Rect &rBlob, const Size &size);
1232
1233	/** @brief Get rectangle coordinates in original image system from rectangle in blob coordinates.
1234	* @param rBlob rect in blob coordinates.
1235	* @param rImg result rect in image coordinates.
1236	* @param size original input image size.
1237	*/
1238	CV_WRAP void blobRectsToImageRects(const std::vector<Rect> &rBlob, CV_OUT std::vector<Rect>& rImg, const Size& size);
1239	};
1240
1241	/** @brief Creates 4-dimensional blob from image with given params.
1242	*
1243	* @details This function is an extension of @ref blobFromImage to meet more image preprocess needs.
1244	* Given input image and preprocessing parameters, and function outputs the blob.
1245	*
1246	* @param image input image (all with 1-, 3- or 4-channels).
1247	* @param param struct of Image2BlobParams, contains all parameters needed by processing of image to blob.
1248	* @return 4-dimensional Mat.
1249	*/
1250	CV_EXPORTS_W Mat blobFromImageWithParams(InputArray image, const Image2BlobParams& param = Image2BlobParams());
1251
1252	/** @overload */
1253	CV_EXPORTS_W void blobFromImageWithParams(InputArray image, OutputArray blob, const Image2BlobParams& param = Image2BlobParams());
1254
1255	/** @brief Creates 4-dimensional blob from series of images with given params.
1256	*
1257	* @details This function is an extension of @ref blobFromImages to meet more image preprocess needs.
1258	* Given input image and preprocessing parameters, and function outputs the blob.
1259	*
1260	* @param images input image (all with 1-, 3- or 4-channels).
1261	* @param param struct of Image2BlobParams, contains all parameters needed by processing of image to blob.
1262	* @returns 4-dimensional Mat.
1263	*/
1264	CV_EXPORTS_W Mat blobFromImagesWithParams(InputArrayOfArrays images, const Image2BlobParams& param = Image2BlobParams());
1265
1266	/** @overload */
1267	CV_EXPORTS_W void blobFromImagesWithParams(InputArrayOfArrays images, OutputArray blob, const Image2BlobParams& param = Image2BlobParams());
1268
1269	/** @brief Parse a 4D blob and output the images it contains as 2D arrays through a simpler data structure
1270	* (std::vector<cv::Mat>).
1271	* @param[in] blob_ 4 dimensional array (images, channels, height, width) in floating point precision (CV_32F) from
1272	* which you would like to extract the images.
1273	* @param[out] images_ array of 2D Mat containing the images extracted from the blob in floating point precision
1274	* (CV_32F). They are non normalized neither mean added. The number of returned images equals the first dimension
1275	* of the blob (batch size). Every image has a number of channels equals to the second dimension of the blob (depth).
1276	*/
1277	CV_EXPORTS_W void imagesFromBlob(const cv::Mat& blob_, OutputArrayOfArrays images_);
1278
1279	/** @brief Convert all weights of Caffe network to half precision floating point.
1280	* @param src Path to origin model from Caffe framework contains single
1281	* precision floating point weights (usually has `.caffemodel` extension).
1282	* @param dst Path to destination model with updated weights.
1283	* @param layersTypes Set of layers types which parameters will be converted.
1284	* By default, converts only Convolutional and Fully-Connected layers'
1285	* weights.
1286	*
1287	* @note Shrinked model has no origin float32 weights so it can't be used
1288	* in origin Caffe framework anymore. However the structure of data
1289	* is taken from NVidia's Caffe fork: https://github.com/NVIDIA/caffe.
1290	* So the resulting model may be used there.
1291	*/
1292	CV_EXPORTS_W void shrinkCaffeModel(const String& src, const String& dst,
1293	const std::vector<String>& layersTypes = std::vector<String>());
1294
1295	/** @brief Create a text representation for a binary network stored in protocol buffer format.
1296	* @param[in] model A path to binary network.
1297	* @param[in] output A path to output text file to be created.
1298	*
1299	* @note To reduce output file size, trained weights are not included.
1300	*/
1301	CV_EXPORTS_W void writeTextGraph(const String& model, const String& output);
1302
1303	/** @brief Performs non maximum suppression given boxes and corresponding scores.
1304
1305	* @param bboxes a set of bounding boxes to apply NMS.
1306	* @param scores a set of corresponding confidences.
1307	* @param score_threshold a threshold used to filter boxes by score.
1308	* @param nms_threshold a threshold used in non maximum suppression.
1309	* @param indices the kept indices of bboxes after NMS.
1310	* @param eta a coefficient in adaptive threshold formula: \f$nms\_threshold_{i+1}=eta\cdot nms\_threshold_i\f$.
1311	* @param top_k if `>0`, keep at most @p top_k picked indices.
1312	*/
1313	CV_EXPORTS void NMSBoxes(const std::vector<Rect>& bboxes, const std::vector<float>& scores,
1314	const float score_threshold, const float nms_threshold,
1315	CV_OUT std::vector<int>& indices,
1316	const float eta = 1.f, const int top_k = 0);
1317
1318	CV_EXPORTS_W void NMSBoxes(const std::vector<Rect2d>& bboxes, const std::vector<float>& scores,
1319	const float score_threshold, const float nms_threshold,
1320	CV_OUT std::vector<int>& indices,
1321	const float eta = 1.f, const int top_k = 0);
1322
1323	CV_EXPORTS_AS(NMSBoxesRotated) void NMSBoxes(const std::vector<RotatedRect>& bboxes, const std::vector<float>& scores,
1324	const float score_threshold, const float nms_threshold,
1325	CV_OUT std::vector<int>& indices,
1326	const float eta = 1.f, const int top_k = 0);
1327
1328	/** @brief Performs batched non maximum suppression on given boxes and corresponding scores across different classes.
1329
1330	* @param bboxes a set of bounding boxes to apply NMS.
1331	* @param scores a set of corresponding confidences.
1332	* @param class_ids a set of corresponding class ids. Ids are integer and usually start from 0.
1333	* @param score_threshold a threshold used to filter boxes by score.
1334	* @param nms_threshold a threshold used in non maximum suppression.
1335	* @param indices the kept indices of bboxes after NMS.
1336	* @param eta a coefficient in adaptive threshold formula: \f$nms\_threshold_{i+1}=eta\cdot nms\_threshold_i\f$.
1337	* @param top_k if `>0`, keep at most @p top_k picked indices.
1338	*/
1339	CV_EXPORTS void NMSBoxesBatched(const std::vector<Rect>& bboxes, const std::vector<float>& scores, const std::vector<int>& class_ids,
1340	const float score_threshold, const float nms_threshold,
1341	CV_OUT std::vector<int>& indices,
1342	const float eta = 1.f, const int top_k = 0);
1343
1344	CV_EXPORTS_W void NMSBoxesBatched(const std::vector<Rect2d>& bboxes, const std::vector<float>& scores, const std::vector<int>& class_ids,
1345	const float score_threshold, const float nms_threshold,
1346	CV_OUT std::vector<int>& indices,
1347	const float eta = 1.f, const int top_k = 0);
1348
1349	/**
1350	* @brief Enum of Soft NMS methods.
1351	* @see softNMSBoxes
1352	*/
1353	enum class SoftNMSMethod
1354	{
1355	SOFTNMS_LINEAR = 1,
1356	SOFTNMS_GAUSSIAN = 2
1357	};
1358
1359	/** @brief Performs soft non maximum suppression given boxes and corresponding scores.
1360	* Reference: https://arxiv.org/abs/1704.04503
1361	* @param bboxes a set of bounding boxes to apply Soft NMS.
1362	* @param scores a set of corresponding confidences.
1363	* @param updated_scores a set of corresponding updated confidences.
1364	* @param score_threshold a threshold used to filter boxes by score.
1365	* @param nms_threshold a threshold used in non maximum suppression.
1366	* @param indices the kept indices of bboxes after NMS.
1367	* @param top_k keep at most @p top_k picked indices.
1368	* @param sigma parameter of Gaussian weighting.
1369	* @param method Gaussian or linear.
1370	* @see SoftNMSMethod
1371	*/
1372	CV_EXPORTS_W void softNMSBoxes(const std::vector<Rect>& bboxes,
1373	const std::vector<float>& scores,
1374	CV_OUT std::vector<float>& updated_scores,
1375	const float score_threshold,
1376	const float nms_threshold,
1377	CV_OUT std::vector<int>& indices,
1378	size_t top_k = 0,
1379	const float sigma = 0.5,
1380	SoftNMSMethod method = SoftNMSMethod::SOFTNMS_GAUSSIAN);
1381
1382
1383	/** @brief This class is presented high-level API for neural networks.
1384	*
1385	* Model allows to set params for preprocessing input image.
1386	* Model creates net from file with trained weights and config,
1387	* sets preprocessing input and runs forward pass.
1388	*/
1389	class CV_EXPORTS_W_SIMPLE Model
1390	{
1391	public:
1392	CV_DEPRECATED_EXTERNAL // avoid using in C++ code, will be moved to "protected" (need to fix bindings first)
1393	Model();
1394
1395	Model(const Model&) = default;
1396	Model(Model&&) = default;
1397	Model& operator=(const Model&) = default;
1398	Model& operator=(Model&&) = default;
1399
1400	/**
1401	* @brief Create model from deep learning network represented in one of the supported formats.
1402	* An order of @p model and @p config arguments does not matter.
1403	* @param[in] model Binary file contains trained weights.
1404	* @param[in] config Text file contains network configuration.
1405	*/
1406	CV_WRAP Model(const String& model, const String& config = "");
1407
1408	/**
1409	* @brief Create model from deep learning network.
1410	* @param[in] network Net object.
1411	*/
1412	CV_WRAP Model(const Net& network);
1413
1414	/** @brief Set input size for frame.
1415	* @param[in] size New input size.
1416	* @note If shape of the new blob less than 0, then frame size not change.
1417	*/
1418	CV_WRAP Model& setInputSize(const Size& size);
1419
1420	/** @overload
1421	* @param[in] width New input width.
1422	* @param[in] height New input height.
1423	*/
1424	CV_WRAP inline
1425	Model& setInputSize(int width, int height) { return setInputSize(Size(width, height)); }
1426
1427	/** @brief Set mean value for frame.
1428	* @param[in] mean Scalar with mean values which are subtracted from channels.
1429	*/
1430	CV_WRAP Model& setInputMean(const Scalar& mean);
1431
1432	/** @brief Set scalefactor value for frame.
1433	* @param[in] scale Multiplier for frame values.
1434	*/
1435	CV_WRAP Model& setInputScale(const Scalar& scale);
1436
1437	/** @brief Set flag crop for frame.
1438	* @param[in] crop Flag which indicates whether image will be cropped after resize or not.
1439	*/
1440	CV_WRAP Model& setInputCrop(bool crop);
1441
1442	/** @brief Set flag swapRB for frame.
1443	* @param[in] swapRB Flag which indicates that swap first and last channels.
1444	*/
1445	CV_WRAP Model& setInputSwapRB(bool swapRB);
1446
1447	/** @brief Set preprocessing parameters for frame.
1448	* @param[in] size New input size.
1449	* @param[in] mean Scalar with mean values which are subtracted from channels.
1450	* @param[in] scale Multiplier for frame values.
1451	* @param[in] swapRB Flag which indicates that swap first and last channels.
1452	* @param[in] crop Flag which indicates whether image will be cropped after resize or not.
1453	* blob(n, c, y, x) = scale * resize( frame(y, x, c) ) - mean(c) )
1454	*/
1455	CV_WRAP void setInputParams(double scale = 1.0, const Size& size = Size(),
1456	const Scalar& mean = Scalar(), bool swapRB = false, bool crop = false);
1457
1458	/** @brief Given the @p input frame, create input blob, run net and return the output @p blobs.
1459	* @param[in] frame The input image.
1460	* @param[out] outs Allocated output blobs, which will store results of the computation.
1461	*/
1462	CV_WRAP void predict(InputArray frame, OutputArrayOfArrays outs) const;
1463
1464
1465	// ============================== Net proxy methods ==============================
1466	// Never expose methods with network implementation details, like:
1467	// - addLayer, addLayerToPrev, connect, setInputsNames, setInputShape, setParam, getParam
1468	// - getLayer*, getUnconnectedOutLayers, getUnconnectedOutLayersNames, getLayersShapes
1469	// - forward* methods, setInput
1470
1471	/// @sa Net::setPreferableBackend
1472	CV_WRAP Model& setPreferableBackend(dnn::Backend backendId);
1473	/// @sa Net::setPreferableTarget
1474	CV_WRAP Model& setPreferableTarget(dnn::Target targetId);
1475
1476	/// @sa Net::enableWinograd
1477	CV_WRAP Model& enableWinograd(bool useWinograd);
1478
1479	CV_DEPRECATED_EXTERNAL
1480	operator Net&() const { return getNetwork_(); }
1481
1482	//protected: - internal/tests usage only
1483	Net& getNetwork_() const;
1484	inline Net& getNetwork_() { return const_cast<const Model*>(this)->getNetwork_(); }
1485
1486	struct Impl;
1487	inline Impl* getImpl() const { return impl.get(); }
1488	inline Impl& getImplRef() const { CV_DbgAssert(impl); return *impl.get(); }
1489	protected:
1490	Ptr<Impl> impl;
1491	};
1492
1493	/** @brief This class represents high-level API for classification models.
1494	*
1495	* ClassificationModel allows to set params for preprocessing input image.
1496	* ClassificationModel creates net from file with trained weights and config,
1497	* sets preprocessing input, runs forward pass and return top-1 prediction.
1498	*/
1499	class CV_EXPORTS_W_SIMPLE ClassificationModel : public Model
1500	{
1501	public:
1502	CV_DEPRECATED_EXTERNAL // avoid using in C++ code, will be moved to "protected" (need to fix bindings first)
1503	ClassificationModel();
1504
1505	/**
1506	* @brief Create classification model from network represented in one of the supported formats.
1507	* An order of @p model and @p config arguments does not matter.
1508	* @param[in] model Binary file contains trained weights.
1509	* @param[in] config Text file contains network configuration.
1510	*/
1511	CV_WRAP ClassificationModel(const String& model, const String& config = "");
1512
1513	/**
1514	* @brief Create model from deep learning network.
1515	* @param[in] network Net object.
1516	*/
1517	CV_WRAP ClassificationModel(const Net& network);
1518
1519	/**
1520	* @brief Set enable/disable softmax post processing option.
1521	*
1522	* If this option is true, softmax is applied after forward inference within the classify() function
1523	* to convert the confidences range to [0.0-1.0].
1524	* This function allows you to toggle this behavior.
1525	* Please turn true when not contain softmax layer in model.
1526	* @param[in] enable Set enable softmax post processing within the classify() function.
1527	*/
1528	CV_WRAP ClassificationModel& setEnableSoftmaxPostProcessing(bool enable);
1529
1530	/**
1531	* @brief Get enable/disable softmax post processing option.
1532	*
1533	* This option defaults to false, softmax post processing is not applied within the classify() function.
1534	*/
1535	CV_WRAP bool getEnableSoftmaxPostProcessing() const;
1536
1537	/** @brief Given the @p input frame, create input blob, run net and return top-1 prediction.
1538	* @param[in] frame The input image.
1539	*/
1540	std::pair<int, float> classify(InputArray frame);
1541
1542	/** @overload */
1543	CV_WRAP void classify(InputArray frame, CV_OUT int& classId, CV_OUT float& conf);
1544	};
1545
1546	/** @brief This class represents high-level API for keypoints models
1547	*
1548	* KeypointsModel allows to set params for preprocessing input image.
1549	* KeypointsModel creates net from file with trained weights and config,
1550	* sets preprocessing input, runs forward pass and returns the x and y coordinates of each detected keypoint
1551	*/
1552	class CV_EXPORTS_W_SIMPLE KeypointsModel: public Model
1553	{
1554	public:
1555	/**
1556	* @brief Create keypoints model from network represented in one of the supported formats.
1557	* An order of @p model and @p config arguments does not matter.
1558	* @param[in] model Binary file contains trained weights.
1559	* @param[in] config Text file contains network configuration.
1560	*/
1561	CV_WRAP KeypointsModel(const String& model, const String& config = "");
1562
1563	/**
1564	* @brief Create model from deep learning network.
1565	* @param[in] network Net object.
1566	*/
1567	CV_WRAP KeypointsModel(const Net& network);
1568
1569	/** @brief Given the @p input frame, create input blob, run net
1570	* @param[in] frame The input image.
1571	* @param thresh minimum confidence threshold to select a keypoint
1572	* @returns a vector holding the x and y coordinates of each detected keypoint
1573	*
1574	*/
1575	CV_WRAP std::vector<Point2f> estimate(InputArray frame, float thresh=0.5);
1576	};
1577
1578	/** @brief This class represents high-level API for segmentation models
1579	*
1580	* SegmentationModel allows to set params for preprocessing input image.
1581	* SegmentationModel creates net from file with trained weights and config,
1582	* sets preprocessing input, runs forward pass and returns the class prediction for each pixel.
1583	*/
1584	class CV_EXPORTS_W_SIMPLE SegmentationModel: public Model
1585	{
1586	public:
1587	/**
1588	* @brief Create segmentation model from network represented in one of the supported formats.
1589	* An order of @p model and @p config arguments does not matter.
1590	* @param[in] model Binary file contains trained weights.
1591	* @param[in] config Text file contains network configuration.
1592	*/
1593	CV_WRAP SegmentationModel(const String& model, const String& config = "");
1594
1595	/**
1596	* @brief Create model from deep learning network.
1597	* @param[in] network Net object.
1598	*/
1599	CV_WRAP SegmentationModel(const Net& network);
1600
1601	/** @brief Given the @p input frame, create input blob, run net
1602	* @param[in] frame The input image.
1603	* @param[out] mask Allocated class prediction for each pixel
1604	*/
1605	CV_WRAP void segment(InputArray frame, OutputArray mask);
1606	};
1607
1608	/** @brief This class represents high-level API for object detection networks.
1609	*
1610	* DetectionModel allows to set params for preprocessing input image.
1611	* DetectionModel creates net from file with trained weights and config,
1612	* sets preprocessing input, runs forward pass and return result detections.
1613	* For DetectionModel SSD, Faster R-CNN, YOLO topologies are supported.
1614	*/
1615	class CV_EXPORTS_W_SIMPLE DetectionModel : public Model
1616	{
1617	public:
1618	/**
1619	* @brief Create detection model from network represented in one of the supported formats.
1620	* An order of @p model and @p config arguments does not matter.
1621	* @param[in] model Binary file contains trained weights.
1622	* @param[in] config Text file contains network configuration.
1623	*/
1624	CV_WRAP DetectionModel(const String& model, const String& config = "");
1625
1626	/**
1627	* @brief Create model from deep learning network.
1628	* @param[in] network Net object.
1629	*/
1630	CV_WRAP DetectionModel(const Net& network);
1631
1632	CV_DEPRECATED_EXTERNAL // avoid using in C++ code (need to fix bindings first)
1633	DetectionModel();
1634
1635	/**
1636	* @brief nmsAcrossClasses defaults to false,
1637	* such that when non max suppression is used during the detect() function, it will do so per-class.
1638	* This function allows you to toggle this behaviour.
1639	* @param[in] value The new value for nmsAcrossClasses
1640	*/
1641	CV_WRAP DetectionModel& setNmsAcrossClasses(bool value);
1642
1643	/**
1644	* @brief Getter for nmsAcrossClasses. This variable defaults to false,
1645	* such that when non max suppression is used during the detect() function, it will do so only per-class
1646	*/
1647	CV_WRAP bool getNmsAcrossClasses();
1648
1649	/** @brief Given the @p input frame, create input blob, run net and return result detections.
1650	* @param[in] frame The input image.
1651	* @param[out] classIds Class indexes in result detection.
1652	* @param[out] confidences A set of corresponding confidences.
1653	* @param[out] boxes A set of bounding boxes.
1654	* @param[in] confThreshold A threshold used to filter boxes by confidences.
1655	* @param[in] nmsThreshold A threshold used in non maximum suppression.
1656	*/
1657	CV_WRAP void detect(InputArray frame, CV_OUT std::vector<int>& classIds,
1658	CV_OUT std::vector<float>& confidences, CV_OUT std::vector<Rect>& boxes,
1659	float confThreshold = 0.5f, float nmsThreshold = 0.0f);
1660	};
1661
1662
1663	/** @brief This class represents high-level API for text recognition networks.
1664	*
1665	* TextRecognitionModel allows to set params for preprocessing input image.
1666	* TextRecognitionModel creates net from file with trained weights and config,
1667	* sets preprocessing input, runs forward pass and return recognition result.
1668	* For TextRecognitionModel, CRNN-CTC is supported.
1669	*/
1670	class CV_EXPORTS_W_SIMPLE TextRecognitionModel : public Model
1671	{
1672	public:
1673	CV_DEPRECATED_EXTERNAL // avoid using in C++ code, will be moved to "protected" (need to fix bindings first)
1674	TextRecognitionModel();
1675
1676	/**
1677	* @brief Create Text Recognition model from deep learning network
1678	* Call setDecodeType() and setVocabulary() after constructor to initialize the decoding method
1679	* @param[in] network Net object
1680	*/
1681	CV_WRAP TextRecognitionModel(const Net& network);
1682
1683	/**
1684	* @brief Create text recognition model from network represented in one of the supported formats
1685	* Call setDecodeType() and setVocabulary() after constructor to initialize the decoding method
1686	* @param[in] model Binary file contains trained weights
1687	* @param[in] config Text file contains network configuration
1688	*/
1689	CV_WRAP inline
1690	TextRecognitionModel(const std::string& model, const std::string& config = "")
1691	: TextRecognitionModel(readNet(model, config)) { /* nothing */ }
1692
1693	/**
1694	* @brief Set the decoding method of translating the network output into string
1695	* @param[in] decodeType The decoding method of translating the network output into string, currently supported type:
1696	* - `"CTC-greedy"` greedy decoding for the output of CTC-based methods
1697	* - `"CTC-prefix-beam-search"` Prefix beam search decoding for the output of CTC-based methods
1698	*/
1699	CV_WRAP
1700	TextRecognitionModel& setDecodeType(const std::string& decodeType);
1701
1702	/**
1703	* @brief Get the decoding method
1704	* @return the decoding method
1705	*/
1706	CV_WRAP
1707	const std::string& getDecodeType() const;
1708
1709	/**
1710	* @brief Set the decoding method options for `"CTC-prefix-beam-search"` decode usage
1711	* @param[in] beamSize Beam size for search
1712	* @param[in] vocPruneSize Parameter to optimize big vocabulary search,
1713	* only take top @p vocPruneSize tokens in each search step, @p vocPruneSize <= 0 stands for disable this prune.
1714	*/
1715	CV_WRAP
1716	TextRecognitionModel& setDecodeOptsCTCPrefixBeamSearch(int beamSize, int vocPruneSize = 0);
1717
1718	/**
1719	* @brief Set the vocabulary for recognition.
1720	* @param[in] vocabulary the associated vocabulary of the network.
1721	*/
1722	CV_WRAP
1723	TextRecognitionModel& setVocabulary(const std::vector<std::string>& vocabulary);
1724
1725	/**
1726	* @brief Get the vocabulary for recognition.
1727	* @return vocabulary the associated vocabulary
1728	*/
1729	CV_WRAP
1730	const std::vector<std::string>& getVocabulary() const;
1731
1732	/**
1733	* @brief Given the @p input frame, create input blob, run net and return recognition result
1734	* @param[in] frame The input image
1735	* @return The text recognition result
1736	*/
1737	CV_WRAP
1738	std::string recognize(InputArray frame) const;
1739
1740	/**
1741	* @brief Given the @p input frame, create input blob, run net and return recognition result
1742	* @param[in] frame The input image
1743	* @param[in] roiRects List of text detection regions of interest (cv::Rect, CV_32SC4). ROIs is be cropped as the network inputs
1744	* @param[out] results A set of text recognition results.
1745	*/
1746	CV_WRAP
1747	void recognize(InputArray frame, InputArrayOfArrays roiRects, CV_OUT std::vector<std::string>& results) const;
1748	};
1749
1750
1751	/** @brief Base class for text detection networks
1752	*/
1753	class CV_EXPORTS_W_SIMPLE TextDetectionModel : public Model
1754	{
1755	protected:
1756	CV_DEPRECATED_EXTERNAL // avoid using in C++ code, will be moved to "protected" (need to fix bindings first)
1757	TextDetectionModel();
1758
1759	public:
1760
1761	/** @brief Performs detection
1762	*
1763	* Given the input @p frame, prepare network input, run network inference, post-process network output and return result detections.
1764	*
1765	* Each result is quadrangle's 4 points in this order:
1766	* - bottom-left
1767	* - top-left
1768	* - top-right
1769	* - bottom-right
1770	*
1771	* Use cv::getPerspectiveTransform function to retrieve image region without perspective transformations.
1772	*
1773	* @note If DL model doesn't support that kind of output then result may be derived from detectTextRectangles() output.
1774	*
1775	* @param[in] frame The input image
1776	* @param[out] detections array with detections' quadrangles (4 points per result)
1777	* @param[out] confidences array with detection confidences
1778	*/
1779	CV_WRAP
1780	void detect(
1781	InputArray frame,
1782	CV_OUT std::vector< std::vector<Point> >& detections,
1783	CV_OUT std::vector<float>& confidences
1784	) const;
1785
1786	/** @overload */
1787	CV_WRAP
1788	void detect(
1789	InputArray frame,
1790	CV_OUT std::vector< std::vector<Point> >& detections
1791	) const;
1792
1793	/** @brief Performs detection
1794	*
1795	* Given the input @p frame, prepare network input, run network inference, post-process network output and return result detections.
1796	*
1797	* Each result is rotated rectangle.
1798	*
1799	* @note Result may be inaccurate in case of strong perspective transformations.
1800	*
1801	* @param[in] frame the input image
1802	* @param[out] detections array with detections' RotationRect results
1803	* @param[out] confidences array with detection confidences
1804	*/
1805	CV_WRAP
1806	void detectTextRectangles(
1807	InputArray frame,
1808	CV_OUT std::vector<cv::RotatedRect>& detections,
1809	CV_OUT std::vector<float>& confidences
1810	) const;
1811
1812	/** @overload */
1813	CV_WRAP
1814	void detectTextRectangles(
1815	InputArray frame,
1816	CV_OUT std::vector<cv::RotatedRect>& detections
1817	) const;
1818	};
1819
1820	/** @brief This class represents high-level API for text detection DL networks compatible with EAST model.
1821	*
1822	* Configurable parameters:
1823	* - (float) confThreshold - used to filter boxes by confidences, default: 0.5f
1824	* - (float) nmsThreshold - used in non maximum suppression, default: 0.0f
1825	*/
1826	class CV_EXPORTS_W_SIMPLE TextDetectionModel_EAST : public TextDetectionModel
1827	{
1828	public:
1829	CV_DEPRECATED_EXTERNAL // avoid using in C++ code, will be moved to "protected" (need to fix bindings first)
1830	TextDetectionModel_EAST();
1831
1832	/**
1833	* @brief Create text detection algorithm from deep learning network
1834	* @param[in] network Net object
1835	*/
1836	CV_WRAP TextDetectionModel_EAST(const Net& network);
1837
1838	/**
1839	* @brief Create text detection model from network represented in one of the supported formats.
1840	* An order of @p model and @p config arguments does not matter.
1841	* @param[in] model Binary file contains trained weights.
1842	* @param[in] config Text file contains network configuration.
1843	*/
1844	CV_WRAP inline
1845	TextDetectionModel_EAST(const std::string& model, const std::string& config = "")
1846	: TextDetectionModel_EAST(readNet(model, config)) { /* nothing */ }
1847
1848	/**
1849	* @brief Set the detection confidence threshold
1850	* @param[in] confThreshold A threshold used to filter boxes by confidences
1851	*/
1852	CV_WRAP
1853	TextDetectionModel_EAST& setConfidenceThreshold(float confThreshold);
1854
1855	/**
1856	* @brief Get the detection confidence threshold
1857	*/
1858	CV_WRAP
1859	float getConfidenceThreshold() const;
1860
1861	/**
1862	* @brief Set the detection NMS filter threshold
1863	* @param[in] nmsThreshold A threshold used in non maximum suppression
1864	*/
1865	CV_WRAP
1866	TextDetectionModel_EAST& setNMSThreshold(float nmsThreshold);
1867
1868	/**
1869	* @brief Get the detection confidence threshold
1870	*/
1871	CV_WRAP
1872	float getNMSThreshold() const;
1873	};
1874
1875	/** @brief This class represents high-level API for text detection DL networks compatible with DB model.
1876	*
1877	* Related publications: @cite liao2020real
1878	* Paper: https://arxiv.org/abs/1911.08947
1879	* For more information about the hyper-parameters setting, please refer to https://github.com/MhLiao/DB
1880	*
1881	* Configurable parameters:
1882	* - (float) binaryThreshold - The threshold of the binary map. It is usually set to 0.3.
1883	* - (float) polygonThreshold - The threshold of text polygons. It is usually set to 0.5, 0.6, and 0.7. Default is 0.5f
1884	* - (double) unclipRatio - The unclip ratio of the detected text region, which determines the output size. It is usually set to 2.0.
1885	* - (int) maxCandidates - The max number of the output results.
1886	*/
1887	class CV_EXPORTS_W_SIMPLE TextDetectionModel_DB : public TextDetectionModel
1888	{
1889	public:
1890	CV_DEPRECATED_EXTERNAL // avoid using in C++ code, will be moved to "protected" (need to fix bindings first)
1891	TextDetectionModel_DB();
1892
1893	/**
1894	* @brief Create text detection algorithm from deep learning network.
1895	* @param[in] network Net object.
1896	*/
1897	CV_WRAP TextDetectionModel_DB(const Net& network);
1898
1899	/**
1900	* @brief Create text detection model from network represented in one of the supported formats.
1901	* An order of @p model and @p config arguments does not matter.
1902	* @param[in] model Binary file contains trained weights.
1903	* @param[in] config Text file contains network configuration.
1904	*/
1905	CV_WRAP inline
1906	TextDetectionModel_DB(const std::string& model, const std::string& config = "")
1907	: TextDetectionModel_DB(readNet(model, config)) { /* nothing */ }
1908
1909	CV_WRAP TextDetectionModel_DB& setBinaryThreshold(float binaryThreshold);
1910	CV_WRAP float getBinaryThreshold() const;
1911
1912	CV_WRAP TextDetectionModel_DB& setPolygonThreshold(float polygonThreshold);
1913	CV_WRAP float getPolygonThreshold() const;
1914
1915	CV_WRAP TextDetectionModel_DB& setUnclipRatio(double unclipRatio);
1916	CV_WRAP double getUnclipRatio() const;
1917
1918	CV_WRAP TextDetectionModel_DB& setMaxCandidates(int maxCandidates);
1919	CV_WRAP int getMaxCandidates() const;
1920	};
1921
1922	//! @}
1923	CV__DNN_INLINE_NS_END
1924	}
1925	}
1926
1927	#include <opencv2/dnn/layer.hpp>
1928	#include <opencv2/dnn/dnn.inl.hpp>
1929
1930	/// @deprecated Include this header directly from application. Automatic inclusion will be removed
1931	#include <opencv2/dnn/utils/inference_engine.hpp>
1932
1933	#endif /* OPENCV_DNN_DNN_HPP */
1934