Deep Neural Network module

Modules
	Partial List of Implemented Layers
	Utilities for New Layers Registration

Classes
class	cv::dnn::BackendNode
	Derivatives of this class encapsulates functions of certain backends. More...
class	cv::dnn::BackendWrapper
	Derivatives of this class wraps cv::Mat for different backends and targets. More...
class	cv::dnn::Dict
	This class implements name-value dictionary, values are instances of DictValue. More...
struct	cv::dnn::DictValue
	This struct stores the scalar value (or array) of one of the following type: double, cv::String or int64. More...
class	cv::dnn::Layer
	This interface class allows to build new Layers - are building blocks of networks. More...
class	cv::dnn::LayerParams
	This class provides all data needed to initialize layer. More...
class	cv::dnn::Net
	This class allows to create and manipulate comprehensive artificial neural networks. More...

Typedefs
typedef std::vector< int >	cv::dnn::MatShape

Enumerations
enum	cv::dnn::Backend {   cv::dnn::DNN_BACKEND_DEFAULT,   cv::dnn::DNN_BACKEND_HALIDE,   cv::dnn::DNN_BACKEND_INFERENCE_ENGINE }
	Enum of computation backends supported by layers. More...
enum	cv::dnn::Target {   cv::dnn::DNN_TARGET_CPU,   cv::dnn::DNN_TARGET_OPENCL }
	Enum of target devices for computations. More...

Functions
Mat	cv::dnn::blobFromImage (InputArray image, double scalefactor=1.0, const Size &size=Size(), const Scalar &mean=Scalar(), bool swapRB=true, bool crop=true)
	Creates 4-dimensional blob from image. Optionally resizes and crops image from center, subtract mean values, scales values by scalefactor, swap Blue and Red channels. More...
void	cv::dnn::blobFromImage (InputArray image, OutputArray blob, double scalefactor=1.0, const Size &size=Size(), const Scalar &mean=Scalar(), bool swapRB=true, bool crop=true)
	Creates 4-dimensional blob from image. More...
Mat	cv::dnn::blobFromImages (InputArrayOfArrays images, double scalefactor=1.0, Size size=Size(), const Scalar &mean=Scalar(), bool swapRB=true, bool crop=true)
	Creates 4-dimensional blob from series of images. Optionally resizes and crops images from center, subtract mean values, scales values by scalefactor, swap Blue and Red channels. More...
void	cv::dnn::blobFromImages (InputArrayOfArrays images, OutputArray blob, double scalefactor=1.0, Size size=Size(), const Scalar &mean=Scalar(), bool swapRB=true, bool crop=true)
	Creates 4-dimensional blob from series of images. More...
void	cv::dnn::imagesFromBlob (const cv::Mat &blob_, OutputArrayOfArrays images_)
	Parse a 4D blob and output the images it contains as 2D arrays through a simpler data structure (std::vector<cv::Mat>). More...
void	cv::dnn::NMSBoxes (const std::vector< Rect > &bboxes, const std::vector< float > &scores, const float score_threshold, const float nms_threshold, std::vector< int > &indices, const float eta=1.f, const int top_k=0)
	Performs non maximum suppression given boxes and corresponding scores. More...
Net	cv::dnn::readNetFromCaffe (const String &prototxt, const String &caffeModel=String())
	Reads a network model stored in Caffe framework's format. More...
Net	cv::dnn::readNetFromCaffe (const char *bufferProto, size_t lenProto, const char *bufferModel=NULL, size_t lenModel=0)
	Reads a network model stored in Caffe model in memory. More...
Net	cv::dnn::readNetFromDarknet (const String &cfgFile, const String &darknetModel=String())
	Reads a network model stored in Darknet model files. More...
Net	cv::dnn::readNetFromTensorflow (const String &model, const String &config=String())
	Reads a network model stored in TensorFlow framework's format. More...
Net	cv::dnn::readNetFromTensorflow (const char *bufferModel, size_t lenModel, const char *bufferConfig=NULL, size_t lenConfig=0)
	Reads a network model stored in TensorFlow framework's format. More...
Net	cv::dnn::readNetFromTorch (const String &model, bool isBinary=true)
	Reads a network model stored in Torch7 framework's format. More...
Mat	cv::dnn::readTorchBlob (const String &filename, bool isBinary=true)
	Loads blob which was serialized as torch.Tensor object of Torch7 framework. More...
void	cv::dnn::shrinkCaffeModel (const String &src, const String &dst, const std::vector< String > &layersTypes=std::vector< String >())
	Convert all weights of Caffe network to half precision floating point. More...

Detailed Description

This module contains:

• API for new layers creation, layers are building bricks of neural networks;
• set of built-in most-useful Layers;
• API to constuct and modify comprehensive neural networks from layers;
• functionality for loading serialized networks models from different frameworks.

Functionality of this module is designed only for forward pass computations (i. e. network testing). A network training is in principle not supported.

Typedef Documentation

MatShape

typedef std::vector<int> cv::dnn::MatShape

Enumeration Type Documentation

Backend

enum cv::dnn::Backend

Enum of computation backends supported by layers.

Enumerator
DNN_BACKEND_DEFAULT  Python: cv.dnn.DNN_BACKEND_DEFAULT
DNN_BACKEND_HALIDE  Python: cv.dnn.DNN_BACKEND_HALIDE
DNN_BACKEND_INFERENCE_ENGINE  Python: cv.dnn.DNN_BACKEND_INFERENCE_ENGINE

Target

enum cv::dnn::Target

Enum of target devices for computations.

Enumerator
DNN_TARGET_CPU  Python: cv.dnn.DNN_TARGET_CPU
DNN_TARGET_OPENCL  Python: cv.dnn.DNN_TARGET_OPENCL

Function Documentation

blobFromImage() [1/2]

Mat cv::dnn::blobFromImage	(	InputArray	image,
		double	scalefactor = 1.0,
		const Size &	size = Size(),
		const Scalar &	mean = Scalar(),
		bool	swapRB = true,
		bool	crop = true
	)

Python:
	retval	=	cv.dnn.blobFromImage(	image[, scalefactor[, size[, mean[, swapRB[, crop]]]]]	)

Creates 4-dimensional blob from image. Optionally resizes and crops image from center, subtract mean values, scales values by scalefactor, swap Blue and Red channels.

Parameters

image	input image (with 1-, 3- or 4-channels).
size	spatial size for output image
mean	scalar with mean values which are subtracted from channels. Values are intended to be in (mean-R, mean-G, mean-B) order if image has BGR ordering and swapRB is true.
scalefactor	multiplier for image values.
swapRB	flag which indicates that swap first and last channels in 3-channel image is necessary.
crop	flag which indicates whether image will be cropped after resize or not

if crop is true, input image is resized so one side after resize is equal to corresponding dimension in size and another one is equal or larger. Then, crop from the center is performed. If crop is false, direct resize without cropping and preserving aspect ratio is performed.

Returns

4-dimansional Mat with NCHW dimensions order.

blobFromImage() [2/2]

void cv::dnn::blobFromImage	(	InputArray	image,
		OutputArray	blob,
		double	scalefactor = 1.0,
		const Size &	size = Size(),
		const Scalar &	mean = Scalar(),
		bool	swapRB = true,
		bool	crop = true
	)

Python:
	retval	=	cv.dnn.blobFromImage(	image[, scalefactor[, size[, mean[, swapRB[, crop]]]]]	)

Creates 4-dimensional blob from image.

This is an overloaded member function, provided for convenience. It differs from the above function only in what argument(s) it accepts.

blobFromImages() [1/2]

Mat cv::dnn::blobFromImages	(	InputArrayOfArrays	images,
		double	scalefactor = 1.0,
		Size	size = Size(),
		const Scalar &	mean = Scalar(),
		bool	swapRB = true,
		bool	crop = true
	)

Python:
	retval	=	cv.dnn.blobFromImages(	images[, scalefactor[, size[, mean[, swapRB[, crop]]]]]	)

Creates 4-dimensional blob from series of images. Optionally resizes and crops images from center, subtract mean values, scales values by scalefactor, swap Blue and Red channels.

Parameters

images	input images (all with 1-, 3- or 4-channels).
size	spatial size for output image
mean	scalar with mean values which are subtracted from channels. Values are intended to be in (mean-R, mean-G, mean-B) order if image has BGR ordering and swapRB is true.
scalefactor	multiplier for images values.
swapRB	flag which indicates that swap first and last channels in 3-channel image is necessary.
crop	flag which indicates whether image will be cropped after resize or not

if crop is true, input image is resized so one side after resize is equal to corresponding dimension in size and another one is equal or larger. Then, crop from the center is performed. If crop is false, direct resize without cropping and preserving aspect ratio is performed.

Returns

4-dimansional Mat with NCHW dimensions order.

blobFromImages() [2/2]

void cv::dnn::blobFromImages	(	InputArrayOfArrays	images,
		OutputArray	blob,
		double	scalefactor = 1.0,
		Size	size = Size(),
		const Scalar &	mean = Scalar(),
		bool	swapRB = true,
		bool	crop = true
	)

Python:
	retval	=	cv.dnn.blobFromImages(	images[, scalefactor[, size[, mean[, swapRB[, crop]]]]]	)

Creates 4-dimensional blob from series of images.

This is an overloaded member function, provided for convenience. It differs from the above function only in what argument(s) it accepts.

imagesFromBlob()

void cv::dnn::imagesFromBlob	(	const cv::Mat &	blob_,
		OutputArrayOfArrays	images_
	)

Python:
	images_	=	cv.dnn.imagesFromBlob(	blob_[, images_]	)

Parse a 4D blob and output the images it contains as 2D arrays through a simpler data structure (std::vector<cv::Mat>).

Parameters

[in]	blob_	4 dimensional array (images, channels, height, width) in floating point precision (CV_32F) from which you would like to extract the images.
[out]	images_	array of 2D Mat containing the images extracted from the blob in floating point precision (CV_32F). They are non normalized neither mean added. The number of returned images equals the first dimension of the blob (batch size). Every image has a number of channels equals to the second dimension of the blob (depth).

NMSBoxes()

void cv::dnn::NMSBoxes	(	const std::vector< Rect > &	bboxes,
		const std::vector< float > &	scores,
		const float	score_threshold,
		const float	nms_threshold,
		std::vector< int > &	indices,
		const float	eta = 1.f,
		const int	top_k = 0
	)

Python:
	indices	=	cv.dnn.NMSBoxes(	bboxes, scores, score_threshold, nms_threshold[, eta[, top_k]]	)

Performs non maximum suppression given boxes and corresponding scores.

Parameters

bboxes	a set of bounding boxes to apply NMS.
scores	a set of corresponding confidences.
score_threshold	a threshold used to filter boxes by score.
nms_threshold	a threshold used in non maximum suppression.
indices	the kept indices of bboxes after NMS.
eta	a coefficient in adaptive threshold formula: \(nms\_threshold_{i+1}=eta\cdot nms\_threshold_i\).
top_k	if >0, keep at most top_k picked indices.

readNetFromCaffe() [1/2]

Net cv::dnn::readNetFromCaffe	(	const String &	prototxt,
		const String &	caffeModel = String()
	)

Python:
	retval	=	cv.dnn.readNetFromCaffe(	prototxt[, caffeModel]	)

Reads a network model stored in Caffe framework's format.

Parameters

prototxt	path to the .prototxt file with text description of the network architecture.
caffeModel	path to the .caffemodel file with learned network.

Returns

Net object.

readNetFromCaffe() [2/2]

Net cv::dnn::readNetFromCaffe	(	const char *	bufferProto,
		size_t	lenProto,
		const char *	bufferModel = NULL,
		size_t	lenModel = 0
	)

Python:
	retval	=	cv.dnn.readNetFromCaffe(	prototxt[, caffeModel]	)

Reads a network model stored in Caffe model in memory.

This is an overloaded member function, provided for convenience. It differs from the above function only in what argument(s) it accepts.

Parameters

bufferProto	buffer containing the content of the .prototxt file
lenProto	length of bufferProto
bufferModel	buffer containing the content of the .caffemodel file
lenModel	length of bufferModel

Returns

Net object.

readNetFromDarknet()

Net cv::dnn::readNetFromDarknet	(	const String &	cfgFile,
		const String &	darknetModel = String()
	)

Python:
	retval	=	cv.dnn.readNetFromDarknet(	cfgFile[, darknetModel]	)

Reads a network model stored in Darknet model files.

Parameters

cfgFile	path to the .cfg file with text description of the network architecture.
darknetModel	path to the .weights file with learned network.

Returns

Network object that ready to do forward, throw an exception in failure cases.

Net object.

readNetFromTensorflow() [1/2]

Net cv::dnn::readNetFromTensorflow	(	const String &	model,
		const String &	config = String()
	)

Python:
	retval	=	cv.dnn.readNetFromTensorflow(	model[, config]	)

Reads a network model stored in TensorFlow framework's format.

Parameters

model	path to the .pb file with binary protobuf description of the network architecture
config	path to the .pbtxt file that contains text graph definition in protobuf format. Resulting Net object is built by text graph using weights from a binary one that let us make it more flexible.

Returns

Net object.

readNetFromTensorflow() [2/2]

Net cv::dnn::readNetFromTensorflow	(	const char *	bufferModel,
		size_t	lenModel,
		const char *	bufferConfig = NULL,
		size_t	lenConfig = 0
	)

Python:
	retval	=	cv.dnn.readNetFromTensorflow(	model[, config]	)

Reads a network model stored in TensorFlow framework's format.

This is an overloaded member function, provided for convenience. It differs from the above function only in what argument(s) it accepts.

Parameters

bufferModel	buffer containing the content of the pb file
lenModel	length of bufferModel
bufferConfig	buffer containing the content of the pbtxt file
lenConfig	length of bufferConfig

readNetFromTorch()

Net cv::dnn::readNetFromTorch	(	const String &	model,
		bool	isBinary = true
	)

Python:
	retval	=	cv.dnn.readNetFromTorch(	model[, isBinary]	)

Reads a network model stored in Torch7 framework's format.

Parameters

model	path to the file, dumped from Torch by using torch.save() function.
isBinary	specifies whether the network was serialized in ascii mode or binary.

Returns

Net object.

Note

Ascii mode of Torch serializer is more preferable, because binary mode extensively use long type of C language, which has various bit-length on different systems.

The loading file must contain serialized nn.Module object with importing network. Try to eliminate a custom objects from serialazing data to avoid importing errors.

List of supported layers (i.e. object instances derived from Torch nn.Module class):

• nn.Sequential
• nn.Parallel
• nn.Concat
• nn.Linear
• nn.SpatialConvolution
• nn.SpatialMaxPooling, nn.SpatialAveragePooling
• nn.ReLU, nn.TanH, nn.Sigmoid
• nn.Reshape
• nn.SoftMax, nn.LogSoftMax

Also some equivalents of these classes from cunn, cudnn, and fbcunn may be successfully imported.

readTorchBlob()

Mat cv::dnn::readTorchBlob	(	const String &	filename,
		bool	isBinary = true
	)

Python:
	retval	=	cv.dnn.readTorchBlob(	filename[, isBinary]	)

Loads blob which was serialized as torch.Tensor object of Torch7 framework.

Warning

This function has the same limitations as readNetFromTorch().

shrinkCaffeModel()

void cv::dnn::shrinkCaffeModel	(	const String &	src,
		const String &	dst,
		const std::vector< String > &	layersTypes = std::vector< String >()
	)

Python:
	None	=	cv.dnn.shrinkCaffeModel(	src, dst[, layersTypes]	)

Convert all weights of Caffe network to half precision floating point.

Parameters

src	Path to origin model from Caffe framework contains single precision floating point weights (usually has .caffemodel extension).
dst	Path to destination model with updated weights.
layersTypes	Set of layers types which parameters will be converted. By default, converts only Convolutional and Fully-Connected layers' weights.

Note

Shrinked model has no origin float32 weights so it can't be used in origin Caffe framework anymore. However the structure of data is taken from NVidia's Caffe fork: https://github.com/NVIDIA/caffe. So the resulting model may be used there.