CN109214505B - Full convolution target detection method of densely connected convolution neural network - Google Patents

Abstract

The invention relates to the field of artificial intelligence, in particular to a full convolution target detection method of a dense connection convolution neural network. The invention provides a full convolution target detection method of a densely connected convolution neural network, aiming at overcoming the defect that the prior method can not detect multi-scale targets more accurately, which is characterized in that the multi-scale feature mapping can be effectively utilized to detect the targets, so that the convolution neural network has higher accuracy in the detection of the targets with different scales in the same image.

Description

A Fully Convolutional Object Detection Method Based on Densely Connected Convolutional Neural Networks

technical field

The invention relates to the field of artificial intelligence, and more particularly, to a fully convolutional target detection method with densely connected convolutional neural networks.

Background technique

Convolutional neural networks are invariant to feature detection. For example, after translating and rotating an object, the convolutional neural network can still recognize them as the same object, but for some objects that occupy a small area in the image, their information will be lost in the process of extracting features by the convolutional neural network. As a result, the target cannot be accurately detected. With the advancement of recent research, it has been found that the use of "multi-scale" feature representation can effectively improve the accuracy of object detection at different scales. Some people have tried to use the image pyramid for multi-scale target detection. The specific method is to first scale an image at multiple scales, and then input images of different scales into the convolutional neural network, but this method requires a lot of The amount of computation and memory is therefore not feasible.

SUMMARY OF THE INVENTION

In order to overcome the inability of existing methods to perform more accurate detection on multi-scale targets, the present invention provides a fully convolutional target detection method with densely connected convolutional neural networks.

In order to achieve the above purpose of the invention, the technical scheme adopted is:

A fully convolutional object detection method based on densely connected convolutional neural networks, which specifically includes the following steps:

Step S1: Build a feature extraction network Densenet. The feature extraction network consists of multiple dense connection blocks and conversion layers. Using dense connection blocks can identify more discriminative visual features in the image. After the input image passes through the feature extraction network, each Features with different semantics and different resolutions output by densely connected blocks;

Step S2: Build a feature pyramid FPN, input the features of each layer retained in step S1 into the FPN, and stack them according to the feature scale to form a bottom-up, scale-increasing low-semantic feature pyramid. The convolution operation is performed through the "parallel path" to obtain higher semantics; at the same time, the convolved feature will be up-sampled to the same scale of the previous layer feature and merged with the previous layer feature, and the feature will continue to be Pass up until the top of the pyramid, and repeat this step until a complete feature pyramid is constructed;

Step S3: Construct a full convolution predictor FCP network. The full convolution predictor FCP is a predictor that can output the target bounding box information and classification probability at the same time, and predicts the feature maps of all scales in the feature pyramid respectively. After the input feature map is passed through a convolutional neural network, a vector of size S*S*(B*5+C) is output as the prediction result, which is equivalent to dividing the original image into S*S grids. Each grid predicts B bounding boxes, and each bounding box contains 5 pieces of information, including the center coordinate offset value of the bounding box (t _x , _ty ), the width and height offset value of the bounding box (t _w , t _{h )} ), and the confidence t ₀ of the predicted bounding box, and the probability of predicting C target classes for each grid;

Step S4: Train the overall network, collect the target image parameters and input them into the network, the parameters of each layer of the network are initialized according to the Xavier method, and use the stochastic gradient descent algorithm of the loss function composed of bounding box coordinate regression and object classification to calculate the loss gradients and fine-tune the parameters in all layers of the entire network using a backpropagation algorithm.

Preferably, the specific steps in the step S1 are as follows:

Step S101 adjusts an existing trained densely connected convolutional neural network model to obtain a preliminary feature extraction network model;

In step S102, the densely connected convolutional neural network is divided into multiple densely connected blocks during the implementation process, and different densely connected blocks are connected through a conversion layer;

Step S103 has multiple convolutional neural network layers in a densely connected block, and the input of each convolutional neural network layer is the superposition of the outputs of all the convolutional neural network layers before it in the same densely connected block; set densely connected The input of the convolutional network of the lth layer in the block is x _l and the output is y _l , then x _l =(x ₁ +y ₁ +...+y _l-1 ), y _l =H(x _l ), where H( .) is defined as the activation function;

Step S104 H(.) is the activation function followed by each layer of the convolutional neural network, here it is a compound operation, indicating that the input x _l first undergoes a BN operation, then a ReLU function, and finally a convolutional layer. Process the output as the entire activation function;

In step S105, due to the different spatial sizes of different dense connection blocks, they are connected to each other through a conversion layer, and the output of a dense connection block above the conversion layer is used as the input, first through a BN operation, and then a convolutional neural network layer. , and finally, through a pooling layer, the spatial size of the feature map is adjusted to match the input of the next dense connection block; here, the spatial size of the feature map after the pooling layer is set to be 1/n times the original size;

In step S106, the dense connection block and the conversion layer are alternately connected for many times, so that the space size of the feature map decreases after each dense connection block, while the number of channels of the feature map increases. The feature map output by a layer of convolutional _neural network is cm;

Step S107 deletes the global average pooling layer and the fully connected classification layer of the existing densely connected convolutional neural network, and uses the feature map output by the last layer of the convolutional neural network of the last densely connected block as the output of the feature extraction network .

Preferably, the specific steps in the step S2 are as follows:

Step S201 FPN consists of a "bottom-up feature pyramid" and a "parallel path". FPN first obtains visual features with different semantics and different scales in each layer from the feature extraction network, and then consists of a "bottom-up" structure. Stacking generates feature pyramids of lower semantic features;

Step S202 takes the feature map output in step S107 as the first input of the FPN, the input feature map uses a convolution layer to adjust the number of channels to a constant d, and uses the feature map adjusted by the number of channels as the lowest of the feature pyramid. Layer feature map, where the feature map of each layer of the feature pyramid is set as D _m ;

Step S203 The "bottom-up path" in the FPN, its main task is to upsample the feature map of the lower layer of the feature pyramid, and the upsampling factor is the reciprocal n of the reduction factor of the pooling layer in the feature extraction network, obtaining The feature map has the same spatial size as the feature map output by the corresponding densely connected block in step S1;

Step S204 the "parallel path" in FPN, which takes the feature map output by each dense connection block in step S1 as input, and then uses a convolution layer to adjust the number of channels of the output feature map to d;

In step S205, through steps S203 and S204, two feature maps with the same spatial size and number of channels are obtained, the corresponding elements of the two feature maps are added, and then a convolutional layer is used to reduce the mixing in the up-sampling process. Therefore, the feature map of the lower layer of the feature pyramid is obtained, and the operations on the input in steps S203 and S204 are respectively denoted as f(.) and g(.), then D _m =g(C _m ), D _k =∫(f(D _k+1 )+g(C _k )), where (0<k<m), ∫ represents the convolution operation in S2.5;

Step S206 repeats step S203, step S204 and step S205, so that the entire feature pyramid is constructed layer by layer from the lowest level of the feature pyramid upward.

Preferably, the specific steps in the step S3 are as follows:

In step S301, a feature pyramid is obtained in step S02, which is characterized in that the feature scale of the feature pyramid increases layer by layer from bottom to top, but the number of channels in each layer remains unchanged, and the spatial size of the feature maps of two adjacent layers is different. The scale factor is n, and a predictor that outputs the target bounding box information and classification probability at the same time is constructed. The predictor will act on the features of each layer of the feature pyramid, so that the network can use feature maps of different scales;

Step S302 is the construction of a predictor that outputs target bounding box information and classification probability, takes a feature map of a certain layer of the feature pyramid as input, and after processing two fully connected layers, outputs a S*S*(B*5+C ) vector as the prediction result, which is equivalent to dividing the original image into S*S grids, and predicting B bounding boxes for each grid, each bounding box contains 5 pieces of information, including the center coordinate deviation of the bounding box. Shift value (t _x , _ty ), the width and height offset value of the bounding box (t _w , _th ), and the confidence t ₀ of the predicted bounding box, and the probability of predicting C target categories for each grid ;

Step S303 Calculation of coordinate values:

x=c _x +σ(t _x )

_y =cy +σ( _ty )

σ(t ₀ )=Pr(object)*IOU(b, object)

where x, y are the actual coordinates of the center of the bounding box in the image, w, h are the width and height of the bounding box, respectively; (c _x , c _y ) are the coordinates of the upper left corner of the grid, p _w , p _h are the input image width and height respectively.

Preferably, the specific steps in the step S4 are as follows:

Step S401 image collection: collect images containing various types of targets in daily life as training images, and each image carries the information about the bounding box and classification of the target in the image obtained after processing;

Step S402 establishes a cost function for each predictor for training, and for the center coordinates of the bounding box, the formula is

As a cost function, for the width and height of the bounding box, the formula

As a cost function, for the predicted class, the formula

表示目标出现在第i个格子中，

表示第i个格子中的第j个边界框对应预测的目标，最终得到如下的代价函数：where λ _coord and λ _noobj are designed to balance the cost function between bounding box and probability costs, and

Indicates that the target appears in the i-th grid,

Represents the predicted target corresponding to the jth bounding box in the ith grid, and finally obtains the following cost function:

Step S403 inputs the marked data collected in step S401 into the network, the parameters of each layer are initialized according to the Xavier method, and the stochastic gradient descent algorithm of the loss function composed of bounding box coordinate regression and object classification is adopted. Calculate the loss gradient and use the backpropagation algorithm to fine-tune the parameters in all layers of the entire network to achieve the purpose of training the network.

Preferably, in the step S1, feature extraction is performed using a network structure in which densely connected blocks and conversion layers are alternately connected, so that a better discriminative feature map in the image can be extracted.

Preferably, the FPN network composed of a densely connected convolutional bottom-up feature pyramid and a parallel path can effectively utilize the feature maps of high-semantic low-scale and high-scale low-semantics to construct a feature map with high semantics Feature pyramid of features, large scale and high localization information.

Compared with the prior art, the beneficial effects of the present invention are:

The present invention provides a full convolution target detection method with densely connected convolutional neural network, which is characterized in that it can effectively use multi-scale feature maps to detect targets, so that the convolutional neural network can detect targets of different scales in the same image. detection has high accuracy.

Description of drawings

FIG. 1 is a flow chart of the present invention.

Detailed ways

The accompanying drawings are for illustrative purposes only, and should not be construed as limitations on this patent;

The present invention will be further elaborated below in conjunction with the accompanying drawings and embodiments.

Example 1

As shown in FIG. 1, the present invention provides a fully convolutional target detection method of densely connected convolutional neural network, which specifically includes the following steps:

Step S1: Build a feature extraction network Densenet. The feature extraction network consists of multiple dense connection blocks and conversion layers. Using dense connection blocks can identify more discriminative visual features in the image. After the input image passes through the feature extraction network, each Features with different semantics and different resolutions output by densely connected blocks;

Step S2: Build a feature pyramid FPN, input the features of each layer retained in step S1 into the FPN, and stack them according to the feature scale to form a bottom-up, scale-increasing low-semantic feature pyramid. The convolution operation is performed through the "parallel path" to obtain higher semantics; at the same time, the convolved feature will be up-sampled to the same scale of the previous layer feature and merged with the previous layer feature, and the feature will continue to be Pass up until the top of the pyramid, and repeat this step until a complete feature pyramid is constructed;

Step S3: Construct a full convolution predictor FCP network. The full convolution predictor FCP is a predictor that can output the target bounding box information and classification probability at the same time, and predicts the feature maps of all scales in the feature pyramid respectively. After the input feature map is passed through a convolutional neural network, a vector of size S*S*(B*5+C) is output as the prediction result, which is equivalent to dividing the original image into S*S grids. Each grid predicts B bounding boxes, and each bounding box contains 5 pieces of information, including the center coordinate offset value of the bounding box (t _x , _ty ), the width and height offset value of the bounding box (t _w , t _{h )} ), and the confidence t ₀ of the predicted bounding box, and the probability of predicting C target classes for each grid;

Step S4: Train the overall network, collect the target image parameters and input them into the network, the parameters of each layer of the network are initialized according to the Xavier method, and use the stochastic gradient descent algorithm of the loss function composed of bounding box coordinate regression and object classification to calculate the loss gradients and fine-tune the parameters in all layers of the entire network using a backpropagation algorithm.

Preferably, the specific steps in the step S1 are as follows:

Step S101 adjusts an existing trained densely connected convolutional neural network model to obtain a preliminary feature extraction network model;

In step S102, the densely connected convolutional neural network is divided into multiple densely connected blocks during the implementation process, and different densely connected blocks are connected through a conversion layer;

Step S103 has multiple convolutional neural network layers in a densely connected block, and the input of each convolutional neural network layer is the superposition of the outputs of all the convolutional neural network layers before it in the same densely connected block; set densely connected The input of the convolutional network of the lth layer in the block is x _l and the output is y _l , then x _l =(x ₁ +y ₁ +...+y _l-1 ), y _l =H(x _l ), where H( .) is defined as the activation function;

Step S104 H(.) is the activation function followed by each layer of the convolutional neural network, here it is a compound operation, indicating that the input x _l first undergoes a BN operation, then a ReLU function, and finally a convolutional layer. Process the output as the entire activation function;

In step S105, due to the different spatial sizes of different dense connection blocks, they are connected to each other through a conversion layer, and the output of a dense connection block above the conversion layer is used as the input, first through a BN operation, and then a convolutional neural network layer. , and finally, through a pooling layer, the spatial size of the feature map is adjusted to match the input of the next dense connection block; here, the spatial size of the feature map after the pooling layer is set to be 1/n times the original size;

In step S106, the dense connection block and the conversion layer are alternately connected for many times, so that the space size of the feature map decreases after each dense connection block, while the number of channels of the feature map increases. The feature map output by a layer of convolutional neural network is C _m ;

Step S107 deletes the global average pooling layer and the fully connected classification layer of the existing densely connected convolutional neural network, and uses the feature map output by the last layer of the convolutional neural network of the last densely connected block as the output of the feature extraction network .

Preferably, the specific steps in the step S2 are as follows:

Step S201 FPN consists of a "bottom-up feature pyramid" and a "parallel path". FPN first obtains visual features with different semantics and different scales in each layer from the feature extraction network, and then consists of a "bottom-up" structure. Stacking generates feature pyramids of lower semantic features;

Step S202 takes the feature map output in step S107 as the first input of the FPN, the input feature map uses a convolution layer to adjust the number of channels to a constant d, and uses the feature map adjusted by the number of channels as the lowest of the feature pyramid. Layer feature map, where the feature map of each layer of the feature pyramid is set as D _m ;

Step S203 The "bottom-up path" in the FPN, its main task is to upsample the feature map of the lower layer of the feature pyramid, and the upsampling factor is the reciprocal n of the reduction factor of the pooling layer in the feature extraction network, obtaining The feature map has the same spatial size as the feature map output by the corresponding densely connected block in step S1;

Step S204 the "parallel path" in FPN, which takes the feature map output by each dense connection block in step S1 as input, and then uses a convolution layer to adjust the number of channels of the output feature map to d;

In step S205, through steps S203 and S204, two feature maps with the same spatial size and number of channels are obtained, the corresponding elements of the two feature maps are added, and then a convolutional layer is used to reduce the mixing in the up-sampling process. Therefore, the feature map of the lower layer of the feature pyramid is obtained, and the operations on the input in steps S203 and S204 are respectively denoted as f(.) and g(.), then D _m =g(C _m ), D _k =∫(f(D _k+1 )+g(C _k )), where (0<k<m), ∫ represents the convolution operation in S2.5;

Step S206 repeats step S203, step S204 and step S205, so that the entire feature pyramid is constructed layer by layer from the lowest level of the feature pyramid upward.

Preferably, the specific steps in the step S3 are as follows:

In step S301, a feature pyramid is obtained in step S02, which is characterized in that the feature scale of the feature pyramid increases layer by layer from bottom to top, but the number of channels in each layer remains unchanged, and the spatial size of the feature maps of two adjacent layers is different. The scale factor is n, and a predictor that outputs the target bounding box information and classification probability at the same time is constructed. The predictor will act on the features of each layer of the feature pyramid, so that the network can use feature maps of different scales;

Step S302 is the construction of a predictor that outputs target bounding box information and classification probability, takes a feature map of a certain layer of the feature pyramid as input, and after processing two fully connected layers, outputs a S*S*(B*5+C ) vector as the prediction result, which is equivalent to dividing the original image into S*S grids, and predicting B bounding boxes for each grid, each bounding box contains 5 pieces of information, including the center coordinate deviation of the bounding box. Shift value (t _x , _ty ), the width and height offset value of the bounding box (t _w , _th ), and the confidence t ₀ of the predicted bounding box, and the probability of predicting C target categories for each grid ;

Step S303 Calculation of coordinate values:

x=c _x +σ(t _x )

_y =cy +σ( _ty )

σ(t ₀ )=Pr(object)*IOU(b,object)

where x, y are the actual coordinates of the center of the bounding box in the image, w, h are the width and height of the bounding box, respectively; (c _x , c _y ) are the coordinates of the upper left corner of the grid, p _w , p _h are the input image width and height respectively.

Preferably, the specific steps in the step S4 are as follows:

Step S401 image collection: collect images containing various types of targets in daily life as training images, and each image carries the information about the bounding box and classification of the target in the image obtained after processing;

Step S402 establishes a cost function for each predictor for training, and for the center coordinates of the bounding box, the formula is

As a cost function, for the width and height of the bounding box, the formula

As a cost function, for the predicted class, the formula

表示目标出现在第i个格子中，

表示第i个格子中的第j个边界框对应预测的目标，最终得到如下的代价函数：where λ _coord and λ _noobj are designed to balance the cost function between bounding box and probability costs, and

Indicates that the target appears in the i-th grid,

Represents the predicted target corresponding to the jth bounding box in the ith grid, and finally obtains the following cost function:

Step S403 inputs the marked data collected in step S401 into the network, the parameters of each layer are initialized according to the Xavier method, and the stochastic gradient descent algorithm of the loss function composed of bounding box coordinate regression and object classification is adopted. Calculate the loss gradient and use the backpropagation algorithm to fine-tune the parameters in all layers of the entire network to achieve the purpose of training the network.

Preferably, in the step S1, feature extraction is performed using a network structure in which densely connected blocks and conversion layers are alternately connected, so that a better discriminative feature map in the image can be extracted.

Preferably, the FPN network composed of a densely connected convolutional bottom-up feature pyramid and a parallel path can effectively utilize the feature maps of high-semantic low-scale and high-scale low-semantics to construct a feature map with high semantics Feature pyramid of features, large scale and high localization information.

Obviously, the above-mentioned embodiments of the present invention are only examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those of ordinary skill in the art, changes or modifications in other different forms can also be made on the basis of the above description. There is no need and cannot be exhaustive of all implementations here. Any modifications, equivalent replacements and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the claims of the present invention.

Claims (7)

1. A full convolution target detection method of a dense connection convolution neural network is characterized by comprising the following steps:

step S1: constructing a feature extraction network, wherein the feature extraction network consists of a plurality of dense connecting blocks and a conversion layer, the dense connecting blocks can be used for identifying visual features which are more discriminative in the image, and the features which are output by each dense connecting block and have different semantics and different resolutions are reserved after the input image passes through the feature extraction network;

step S2: constructing a feature pyramid FPN, inputting each layer of features reserved in the step S1 into the FPN, stacking according to feature scales to form a low-semantic feature pyramid with scales increasing from bottom to top, and performing convolution operation on each layer of features through a parallel path from the lowest layer to obtain higher semantic; simultaneously, the convolved features are sampled to the same scale of the previous layer of features and are combined with the previous layer of features, the features are continuously transmitted upwards until the top of the pyramid, and the step is circulated until a complete feature pyramid is constructed;

step S3: constructing a full convolution predictor FCP network, wherein the full convolution predictor FCP is a predictor capable of simultaneously outputting target boundary box information and classification probability, predicting feature mappings of all scales in a feature pyramid respectively, enabling the input feature mappings to pass through a convolution neural network and then outputting a vector with the size of S (B5 + C) as a prediction result, the function of the predictor is equivalent to dividing an original image into S grids, predicting B boundary boxes for each grid, and each boundary box contains 5 pieces of information and comprises a central coordinate deviation value (t) of the boundary box_x,t_y) Width and height offset value (t) of bounding box_w,t_h) And confidence t of the predicted bounding box₀And also the probability of predicting C object classes for each mesh;

step S4: training the whole network, collecting target image parameters and inputting the target image parameters into the network, initializing the parameters of each layer of network according to an Xavier mode, calculating a loss gradient by adopting a random gradient descent algorithm of a loss function consisting of bounding box coordinate regression and object classification, and finely adjusting the parameters in all layers in the whole network by using a reverse conduction algorithm.

2. The method for detecting the full convolution target of the densely-connected convolutional neural network of claim 1, wherein the specific steps in the step S1 are as follows:

s101, adjusting an existing trained dense connection convolutional neural network model to obtain a primary feature extraction network model;

s102, the dense connection convolutional neural network is divided into a plurality of dense connection blocks in the implementation process, and different dense connection blocks are connected through conversion layers;

step S103, a plurality of convolutional neural network layers are arranged in a dense connection block, and the input of each convolutional neural network layer is the superposition of the outputs of all the convolutional neural network layers before the convolutional neural network layer in the same dense connection block; let the convolutional network input of the l-th layer in the dense connection block be x_lOutput is y_lThen x_l＝(x₁+y₁+…+y_l-1)，y_l＝H(x_l) Where H (.) is defined as the activation function;

step S104H (.) is the activation function followed by each layer of the convolutional neural network, where it is a complex operation representing the input x_lFirst, a BN operation is carried out, then a ReLU function is carried out, and finally a convolution layer is processed to be used as the output of the whole activation function;

s105, because the space sizes of different dense connecting blocks are different, the dense connecting blocks are connected with each other through a conversion layer, the output of more than one dense connecting block of the conversion layer is used as input, the output is firstly subjected to BN operation, then is connected with a convolutional neural network layer, and finally the space size of feature mapping is adjusted to be in accordance with the input of the next dense connecting block through a pooling layer; setting the space size after the characteristic mapping of the pooling layer to be 1/n times of the original space size;

step S106, the dense connecting blocks and the conversion layers are alternately connected for a plurality of times, so that the space size of the feature mapping is reduced after passing through one dense connecting block every time, the number of channels of the feature mapping is increased, and the feature mapping output by the last layer of convolutional neural network of each dense connecting block is set as C_m；

Step S107 deletes the global average pooling layer and the fully-connected classification layer of the existing densely-connected convolutional neural network, and uses the feature mapping output by the last layer of convolutional neural network of the last densely-connected block as the output of the feature extraction network.

3. The method for detecting the full convolution target of the densely-connected convolutional neural network of claim 2, wherein the specific steps in the step S2 are as follows:

step S201FPN is composed of a bottom-up feature pyramid and a parallel path, the FPN firstly obtains visual features of different semantemes and different scales of each layer from a feature extraction network, and then the feature pyramid with lower semantic features is generated by stacking bottom-up structures;

step S202 takes the feature mapping output in step S107 as the first input of FPN, the input feature mapping uses a convolution layer to adjust the channel number to a constant D, and the feature mapping after the channel number adjustment is used as the feature mapping of the lowest layer of the feature pyramid, wherein the feature mapping of each layer of the feature pyramid is set as D_m；

The 'bottom-up path' in the step S203FPN has a main task of upsampling the feature map of the lower layer of the feature pyramid, wherein the factor of the upsampling is the reciprocal n of the reduction factor of the pooling layer in the feature extraction network, and the obtained feature map has the same space size as the feature map output by the corresponding dense connection block in the step S1;

a "parallel path" in step S204FPN, which takes the feature map output from each dense connection block in step S1 as input, and then adjusts the number of channels of the output feature map to d using one convolutional layer;

step S205, through step S203 and step S204, obtains two feature maps with the same spatial size and channel number, adds corresponding elements to the two feature maps, and then reduces aliasing effect in the upsampling process through a convolutional layer, thereby obtaining a feature map with a lower feature pyramid, and records the input operations in step S203 and step S204 as f () and g (), respectively, then D_m＝g(C_m)，D_k＝∫(f(D_k+1)+g(C_k) Wherein k is more than 0 and less than m, and the formulaIllustrating the convolution operation in S205;

step S206 repeats step S203, step S204, and step S205, so that the entire feature pyramid is constructed layer by layer upward from the lowest layer of the feature pyramid.

4. The method for detecting the full convolution target of the densely-connected convolutional neural network of claim 2, wherein the specific steps in the step S3 are as follows:

step S301 is to obtain a feature pyramid in step S02, wherein the feature scale of the feature pyramid is increased from bottom to top layer by layer, but the number of channels in each layer is kept unchanged, the scale factor of the space size of the feature mapping of two adjacent layers is n, a predictor for simultaneously outputting the information of the target boundary box and the classification probability is constructed, and the predictor acts on the features of each layer of the feature pyramid, so that the network can utilize the feature mapping of different scales;

step S302 is to construct a predictor for outputting target bounding box information and classification probability, which takes a feature map of a certain layer of a feature pyramid as input, and after processing of two fully connected layers, outputs a vector of S × S (B × 5+ C) as a prediction result, which is equivalent to dividing an original image into S × S meshes, and predicts B bounding boxes for each mesh, each bounding box containing 5 pieces of information, including a central coordinate offset value (t) of the bounding box_x，t_y) Width and height offset value (t) of bounding box_w，t_h) And confidence t of the predicted bounding box₀And also the probability of predicting C object classes for each mesh;

step S303 calculation of coordinate values:

x＝c_x+σ(t_x)

y＝c_y+σ(t_y)

σ(t₀)＝Pr(object)*IOU(b，object)

wherein x and y are the actual coordinates of the center of the bounding box in the image, and w and h are the width and height of the bounding box respectively; (c)_x，c_y) Is the coordinate of the upper left corner of the grid, p_w，p_hThe width and height of the input image, respectively.

5. The method for detecting the full convolution target of the densely-connected convolutional neural network of claim 1, wherein the specific steps in the step S4 are as follows:

step S401, image acquisition: collecting images containing various targets in daily life as training images, wherein each image is provided with processed information about a boundary frame and classification of the targets in the image;

step S402, establishing a cost function for each pre-measurement for training, and adopting a formula for the center coordinates of the bounding box

As a cost function, for the width and height of the bounding box, a formula is adopted

As a cost function, for the prediction class, a formula is adopted

Wherein λ is_coordAnd λ_noobjIn order to balance the cost function between the cost of the bounding box and the probability

The presentation target appears in the ith cell,

representing the target corresponding to the prediction of the jth bounding box in the ith lattice, and finally obtaining the following cost function:

step S403 inputs the marked data collected in step S401 into the network, initializes the parameters of each layer in the Xavier manner, calculates the loss gradient by using a random gradient descent algorithm of the loss function composed of bounding box coordinate regression and object classification, and finely adjusts the parameters in all layers in the entire network by using a back-propagation algorithm, thereby achieving the purpose of training the network.

6. The method of claim 1, wherein in step S1, the feature extraction is performed by using a network structure in which the densely connected blocks and the conversion layers are alternately connected, so as to extract a feature map with better discriminability in the image.

7. The method for detecting the full-convolution target of the densely-connected convolutional neural network as claimed in claim 1, wherein a FPN network composed of a "bottom-up feature pyramid" and a "parallel path" is used, so that feature mapping of high semantic low-scale and high-scale low semantic can be effectively utilized to construct a feature pyramid with high semantic features, large scale and high positioning information.

Images