CN109784476B - Method for improving DSOD network - Google Patents

Abstract

The invention relates to a method for improving DSOD network. First, the input image is preprocessed, the preprocessed image is input into the DSOD feature extraction sub-network, and the RFB_a network is added after the second transition layer of the feature extraction sub-network. The module extracts features with different receptive fields through Atrous convolution with different sampling steps in the RFB_a network. After the feature extraction sub-network, an Atrous convolution layer with a sampling step of 6 is added, and the features generated by the Atrous convolution layer are input to In the multi-scale prediction layer, the multi-scale prediction layer is input into the loss function, and the IOG penalty term is added to the loss function to prevent the overlapping of similar prediction boxes when predicting dense targets of the same type. At the same time, a warm-up strategy is used to set the learning rate in the training phase, and by setting an appropriate batch sample size, the hardware equipment requirements for training the network are reduced. Compared with the original DSOD algorithm, the present invention has higher detection accuracy, improves the detection ability of small targets, and reduces the hardware equipment requirements for training the network.

Description

A Method for Improving DSOD Network

technical field

The invention relates to the field of computer vision, in particular to a method for improving DSOD network.

Background technique

Object detection is one of the most important research topics in computer vision. The main task is to locate the objects of interest in a given image and accurately determine the specific location of each object. Target detection algorithms based on convolutional neural networks can be divided into two types: target detection algorithms based on region extraction and target detection algorithms based on regression. The target detection algorithm based on region extraction has high detection accuracy, but it needs to extract candidate regions, and the detection speed is difficult to achieve real-time. Regression-based target detection algorithms, such as SSD and DSOD, achieve real-time detection by removing the step of region extraction. However, the DSOD algorithm has the problems of poor detection ability of small targets and high requirements for hardware equipment when training the network.

SUMMARY OF THE INVENTION

In view of this, the purpose of the present invention is to propose a method for improving the DSOD network, which can improve the detection ability of small targets and the detection accuracy of the targets.

The present invention adopts the following scheme to realize: a kind of method for improving DSOD network, comprises the following steps:

Step S1: Obtain an image in the dataset as an input image, and input the input image to the input layer; perform preprocessing on the input image by cropping, mirroring and de-averaging to obtain a preprocessed image, and at the same time use the normalization method to convert the preprocessed image. Convert absolute coordinates in the processed image to relative coordinates;

Step S2: adding the RFB_a network module after the second transition layer of the feature extraction sub-network in the DSOD network; inputting the preprocessed image in step S1 into the feature extraction sub-network in the DSOD network for feature extraction; The feature map of the second transition layer in the feature extraction sub-network in the network is input into the RFB_a network module, and features with different receptive fields are extracted through Atrous dilation convolution with different sampling steps in the RFB_a network module; the The extracted features of different receptive fields are input into the 3×3 convolutional layer to form the first scale prediction layer of the DSOD network;

Step S3: after adding the Atrous convolutional layer with preset sampling step size to the feature extraction sub-network in the DSOD network, and input the feature map of the feature extraction sub-network described in step S2 into the feature extraction sub-network with preset (sampling). The step size is 6) In the Atrous convolutional layer with the sampling step size, it is used to increase the receptive field of the feature map; at the same time, the features generated by the Atrous convolutional layer are input into the multi-scale prediction layer in the DSOD network to form 5 scale prediction layer;

Step S4: Input the features of the first scale prediction layer of the DSOD network described in the step S2 and the 5 scale prediction layers described in the step S3 into the multi-task loss function L adding the IOG penalty term;

Step S5: set the learning rate through the preheating strategy, and use the gradient descent algorithm to optimize the weights of all network layers in the DSOD network; set an appropriate sample size (the present invention is set to 16), in order to reduce the hardware for training the DSOD network equipment requirements.

Further, the image preprocessing in step S1 is specifically:

Step S11: Crop the input image: first randomly select the input image, that is, the length and height of the cropped image, and then randomly select a number among 0.1, 0.3, 0.7, and 0.9 as the Jaccard coefficient threshold, and calculate all the values in the original image through the Jaccard coefficient. The similarity of the ground-truth box and the cropped image;

Step S12: Judging whether the Jaccard coefficient of the real frame or the cropped image is greater than the Jaccard threshold randomly selected in step S11; if at least one real frame and the Jaccard coefficient of the cropped image are greater than the selected Jaccard threshold, and the center coordinates of the real frame are If it falls in the cropped image, the cropped image meets the requirements, otherwise it returns to step S11; the Jaccard coefficient is calculated as follows:

Among them, N represents the number of real boxes in the image, box _i represents the area of the ith real box, box _cut represents the area of the cropped image, and operator ∩ represents the calculation of the overlap area.

Step S13: Perform left and right mirroring processing on the cropped image according to a preset probability T (T=0.5 in the present invention), and adjust the resolution of the mirrored image to 300×300 to obtain a mirrored image.

Step S14: De-average the mirrored image by using a de-average method to obtain a de-averaged image.

Further, the specific content of the step S2 is: first, a 1×1 convolution layer is used on each RFB_a network branch to reduce the number of channels of the feature; on the first branch of the RFB_a network, the step size is 1 and The convolution kernel is a 3×3 convolutional layer, and a 3×3 receptive field feature is obtained; the second branch adopts a 1×3 convolutional layer and an Atrous convolutional layer with a sampling step size of 3 to obtain a 1×7 convolutional layer. Receptive field features; use a 3×1 convolutional layer and an Atrous convolutional layer with a sampling stride of 3 on the third branch to obtain a 7×1 receptive field feature; use a 3×3 convolution on the fourth branch layer and the Atrous convolutional layer with sampling step size of 5 to obtain 11×11 receptive field features; the features extracted from each branch are fused through channel splicing and 1×1 convolutional layers; finally, the fused features are combined with the DSOD network The features produced by the second transition layer are fused by residuals to form the final output features.

Further, the specific method of adding the Atrous convolutional layer with a preset sampling step size described in step S3 is: first, increase the number of output channels C of the feature extraction sub-network in the DSOD network, in order to extract more abundant features. information, and then add the Atrous convolution layer with a certain sampling step size r. The output channel of the Atrous convolution layer is equal to the number of output channels of the original DSOD feature extraction sub-network, so that the Atrous convolution is embedded in the DSOD network; then add 1 A ×1 convolutional layer performs feature fusion.

Further, the multi-task loss function L added to the IOG penalty item described in step S4 is specifically:

Step S41: Calculate the prediction frame g _{iou_max} with the largest area intersection ratio between the prediction frame p output by the DSOD network and all the real frames G of the preset sample, and the formula is as follows:

Among them, g represents the real box, G represents the set of all real boxes, p represents the prediction box, P represents the set of all prediction boxes, box _g represents the area of the real box, and box _p represents the area of the prediction box;

Step S42: Remove the real frame with the largest area of the intersection and ratio in step S41, then calculate the largest IOG penalty item between the predicted frame and the remaining real frame, and use the largest IOG penalty item as the _Liog loss function. The calculation formula is as follows:

Step S43: Weighted fusion of the _Liog loss function, the localization loss function _Lloc and the classification loss function L _conf function to form the final multi-task loss function L, the formula is as follows:

Among them, N represents the number of detected positive samples, α represents the weight of L _loc of the localization loss; the localization loss function L _loc adopts the smooth _L1 loss; the classification loss function L _conf adopts the information cross entropy to calculate; the localization loss function L _loc The calculation is as follows:

指示函数，表示第i个默认框匹配第j个类别为k的真实框，l表示预测框的位置坐标，pos表示正样本的默认框，N表示的正样本的数量；smooth_L1计算如下：in,

Indication function, indicating that the ith default box matches the jth category of the real box, l represents the position coordinates of the predicted box, pos represents the default box of the positive sample, and N represents the number of positive samples; smooth _L1 is calculated as follows:

The classification loss _Lconf is calculated as follows:

表示第i个预测框为类别p的概率，计算如下：Among them, c represents the confidence of each category, Neg represents the negative sample, p represents the category, and 0 represents the category is the background.

Represents the probability that the i-th prediction box is of category p, calculated as follows:

Further, using the warmup strategy to set the learning rate in step S5 is specifically: setting the initial learning rate to 10 ^-5 , and linearly increasing the learning rate to 10 -2 in the first 5 epochs, and in the 75th epoch, the learning rate is linearly increased to 10 ^-2 . epoch, 125th epoch and 175th epoch divide the learning rate by 10, and complete the training at the 200th epoch; the initial value of batch normalization weights is set to 0.5, and the value of bias is set to 0; all volumes The product uses the xavier method for initialization; by improving the training strategy, the training batch sample size is reduced from 128 to 16 to reduce the hardware equipment requirements for training the network.

Compared with the prior art, the present invention has the following beneficial effects:

The invention adds an efficient network structure to the low-level network, extracts more global feature information, and improves the detection ability of small targets; a penalty term is added to the loss function to prevent the overlapping of similar prediction frames when the targets are dense, and the non-maximum The value suppresses the missed detection during post-processing, which improves the detection accuracy of the target. In addition, by improving the training strategy, the hardware equipment requirements for training the network are reduced.

Description of drawings

FIG. 1 is a structural diagram of an embodiment of the present invention.

FIG. 2 is a one-dimensional feature extraction diagram of a convolution layer and an Atrous convolution layer according to an embodiment of the present invention.

FIG. 3 is a network structure diagram of RFB_a according to an embodiment of the present invention.

FIG. 4 is a dense sampling diagram of an embodiment of the present invention.

5 is a comparison diagram of the detection result of the specific embodiment 1 of the embodiment of the present invention and the detection result of the original DSOD.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings and embodiments.

As shown in Figure 1, the present embodiment provides a method for improving a DSOD network, including the following steps:

Step S1: Obtain an image in the dataset as an input image, and input the input image to the input layer; perform preprocessing on the input image by cropping, mirroring and de-averaging to obtain a preprocessed image, and at the same time use the normalization method to convert the preprocessed image. Convert absolute coordinates in the processed image to relative coordinates;

Step S2: adding the RFB_a network module after the second transition layer of the feature extraction sub-network in the DSOD network; inputting the preprocessed image in step S1 into the feature extraction sub-network in the DSOD network for feature extraction; The feature map of the second transition layer in the feature extraction sub-network in the network is input into the RFB_a network module, and features with different receptive fields are extracted through Atrous dilation convolution with different sampling steps in the RFB_a network module; the The extracted features of different receptive fields are input into the 3×3 convolutional layer to form the first scale prediction layer of the DSOD network;

Step S3: after adding the Atrous convolutional layer with preset sampling step size to the feature extraction sub-network in the DSOD network, and input the feature map of the feature extraction sub-network described in step S2 into the feature extraction sub-network with preset (sampling). The step size is 6) In the Atrous convolutional layer with the sampling step size, it is used to increase the receptive field of the feature map; at the same time, the features generated by the Atrous convolutional layer are input into the multi-scale prediction layer in the DSOD network to form 5 scale prediction layer;

Step S4: Input the features of the first scale prediction layer of the DSOD network described in the step S2 and the 5 scale prediction layers described in the step S3 into the multi-task loss function L adding the IOG penalty term;

Step S5: set the learning rate through the preheating strategy, and use the gradient descent algorithm to optimize the weights of all network layers in the DSOD network; set an appropriate sample size (the present invention is set to 16), in order to reduce the hardware for training the DSOD network equipment requirements.

In this embodiment, the image preprocessing in step S1 is specifically:

Step S11: Crop the input image: first randomly select the input image, that is, the length and height of the cropped image, and then randomly select a number among 0.1, 0.3, 0.7, and 0.9 as the Jaccard coefficient threshold, and calculate all the values in the original image through the Jaccard coefficient. The similarity of the ground-truth box and the cropped image;

Step S12: Judging whether the Jaccard coefficient of the real frame or the cropped image is greater than the Jaccard threshold randomly selected in step S11; if at least one real frame and the Jaccard coefficient of the cropped image are greater than the selected Jaccard threshold, and the center coordinates of the real frame are If it falls in the cropped image, the cropped image meets the requirements, otherwise it returns to step S11; the Jaccard coefficient is calculated as follows:

Among them, N represents the number of real boxes in the image, box _i represents the area of the ith real box, box _cut represents the area of the cropped image, and operator ∩ represents the calculation of the overlap area.

Step S13: Perform left and right mirroring processing on the cropped image according to a preset probability T (T=0.5 in the present invention), and adjust the resolution of the mirrored image to 300×300 to obtain a mirrored image.

Step S14: De-average the mirrored image by using a de-average method to obtain a de-averaged image.

In this embodiment, the specific content of step S2 is: first, a 1×1 convolution layer is used on each RFB_a network branch to reduce the number of feature channels; step size is used on the first branch of the RFB_a network is 1 and the convolution kernel is a 3×3 convolutional layer to obtain a 3×3 receptive field feature; the second branch uses a 1×3 convolutional layer and an Atrous convolutional layer with a sampling stride of 3 to obtain 1 ×7 receptive field features; use 3 × 1 convolutional layers and Atrous convolutional layers with sampling stride 3 on the third branch to obtain 7 × 1 receptive field features; use 3 × 1 on the fourth branch 3 convolutional layers and Atrous convolutional layers with a sampling step size of 5 to obtain 11×11 receptive field features; the features extracted from each branch are fused through channel splicing and 1×1 convolutional layers; finally, the fused features and The features generated by the second transition layer of the DSOD network are fused by residuals to form the final output features.

In this embodiment, the specific method for adding the Atrous convolutional layer with a preset sampling step size in step S3 is: first, increase the number of output channels C of the feature extraction sub-network in the DSOD network, in order to extract more Rich feature information, and then add the Atrous convolution layer with a certain sampling step size r, the output channel of the Atrous convolution layer is equal to the number of output channels of the original DSOD feature extraction sub-network, so that the Atrous convolution is embedded in the DSOD network; Then a 1×1 convolutional layer is added for feature fusion.

In this embodiment, the multi-task loss function L added with the IOG penalty item in step S4 is specifically:

Step S41: Calculate the prediction frame g _{iou_max} with the largest area intersection ratio between the prediction frame p output by the DSOD network and all the real frames G of the preset sample, and the formula is as follows:

Among them, g represents the real box, G represents the set of all real boxes, p represents the prediction box, P represents the set of all prediction boxes, box _g represents the area of the real box, and box _p represents the area of the prediction box;

Step S42: Remove the real frame with the largest area of the intersection and ratio in step S41, then calculate the largest IOG penalty item between the predicted frame and the remaining real frame, and use the largest IOG penalty item as the _Liog loss function. The calculation formula is as follows:

Step S43: Weighted fusion of the _Liog loss function, the localization loss function _Lloc and the classification loss function L _conf function to form the final multi-task loss function L, the formula is as follows:

Among them, N represents the number of detected positive samples, α represents the weight of L _loc of the localization loss; the localization loss function L _loc adopts the smooth _L1 loss; the classification loss function L _conf adopts the information cross entropy to calculate; the localization loss function L _loc The calculation is as follows:

指示函数，表示第i个默认框匹配第j个类别为k的真实框，l表示预测框的位置坐标，pos表示正样本的默认框，N表示的正样本的数量；smooth_L1计算如下：in,

Indication function, indicating that the ith default box matches the jth category of the real box, l represents the position coordinates of the predicted box, pos represents the default box of the positive sample, and N represents the number of positive samples; smooth _L1 is calculated as follows:

The classification loss _Lconf is calculated as follows:

表示第i个预测框为类别p的概率，计算如下：Among them, c represents the confidence of each category, Neg represents the negative sample, p represents the category, and 0 represents the category is the background.

Represents the probability that the i-th prediction box is of category p, calculated as follows:

In this embodiment, the use of the warmup strategy to set the learning rate in step S5 is specifically: setting the initial learning rate to 10 ⁻⁵ , and linearly increasing the learning rate to 10 ⁻² in the first 5 epochs. The 75th epoch, the 125th epoch and the 175th epoch divide the learning rate by 10 respectively, and the training is completed at the 200th epoch; the initial value of the batch normalization weight is set to 0.5, and the value of the bias is set to 0; All convolutions are initialized using the xavier method; the training batch size is reduced from 128 to 16 by improving the training strategy to reduce the hardware requirements for training the network.

Preferably, in this embodiment, the input image is preprocessed by cropping, mirroring, and de-averaging; the preprocessed image is input into the DSOD feature extraction sub-network, after the second transition layer of the feature extraction sub-network. In the RFB_a network module, the features with different receptive fields are extracted through Atrous convolution with different sampling steps in the RFB_a network, which provides features with more global information for the detection of small targets. After the feature extraction sub-network, the sampling step is added as 6's Atrous convolutional layer increases the receptive field of the feature map and provides richer semantic information for the subsequent multi-scale prediction layer; the features generated by the Atrous convolutional layer are input into the multi-scale prediction layer, and the multi-scale prediction layer Input into the loss function, and add the IOG penalty term to the loss function to prevent overlapping of similar prediction boxes when predicting dense targets of the same type, so as to avoid missed detection after being processed by non-maximum suppression; The warmup strategy sets the learning rate, which reduces the hardware requirements for training the network through an appropriate batch sample size. The experimental results show that compared with the original DSOD algorithm, the present invention has higher detection accuracy, improves the detection ability of small targets, and reduces the hardware equipment requirements for training the network.

Figure 2 is a one-dimensional feature extraction diagram of the standard convolutional layer and the Atrous convolutional layer. When the sampling stride r is 1, the Atrous convolution is a standard convolution. When the sampling step r is 2, the padding coefficient is 2, and r-1 0s are inserted between the input signals, after the Atrous convolution operation, the 3 input signals generate 5 signal excitations. As can be seen from the figure, the Atrous convolutional layer has the receptive field of increasing the convolution kernel. The use of the two-dimensional Atrous convolution in this embodiment also has the same effect as the one-dimensional Atrous convolution.

Figure 3 shows the network structure of a standard RFB_a. The RFB_a network module is the structure of a multi-branch convolutional network. In different branches, the receptive field features of different sizes extracted by Atrous convolution with different sampling steps are fused by channel splicing to form the effect of dense sampling on the original feature map, as shown in Figure 4. In this embodiment, a ReLU activation function is added to the last Atrous convolution of each branch in the standard RFB_a network structure to extract higher-level features. Meanwhile, in order to ensure the consistency of the DSOD network structure, this embodiment adjusts the batch normalization (BN) and ReLU activation functions in the RFB_a network before the convolution layer.

Example 1, as shown in FIG. 5 , uses the target detection analysis tool to analyze the detection capability of the DSOD and the improved method on the smallest (XS) target. It can be seen from Figure 5 that, except for the table category, the detection accuracy of the improved DSOD object detection method in categories such as airplanes, bicycles, and birds has been improved to varying degrees. In the table category image, the cups and other items placed on the table cause a certain occlusion on the table, which has a great impact on the improved DSOD detection, so the accuracy is lower than the original DSOD algorithm. In general, the improved method of this embodiment has better detection accuracy for small targets.

Embodiment 2, in the PASCAL VOC2007 test set, the improved DSOD and some other typical regression-based target detection algorithms are compared in terms of detection accuracy and detection speed. The main indicators considered are mAP (mean Average Precision) and FPS (Frames Per Second). Among them, * represents the result of testing in the experimental environment of this embodiment. From the data in the table, it can be seen that the improved DSOD model has higher accuracy, and the detection accuracy is improved from 77.4% to 79.0%. Compared with DSSD, the improved DSOD is superior to DSSD in both detection accuracy and detection speed. Since RFBNet300 uses multiple RFB network blocks, more global features can be extracted, and the accuracy is roughly the same as the improved method in this embodiment. However, the computational complexity of RFBNet300 is higher than that of the improved method in this embodiment, and the improved method has better real-time performance.

The above descriptions are only preferred embodiments of the present invention, and all equivalent changes and modifications made according to the scope of the patent application of the present invention shall fall within the scope of the present invention.

Claims (5)

1. a method for improving DSOD network, is characterized in that: comprise the following steps: Step S1: Obtain an image in the dataset as an input image, and input the input image to the input layer; perform preprocessing on the input image by cropping, mirroring and de-averaging to obtain a preprocessed image, and at the same time use the normalization method to convert the preprocessed image. Convert absolute coordinates in the processed image to relative coordinates; Step S2: adding the RFB_a network module after the second transition layer of the feature extraction sub-network in the DSOD network; inputting the preprocessed image in step S1 into the feature extraction sub-network in the DSOD network for feature extraction; The feature map of the second transition layer in the feature extraction sub-network in the network is input into the RFB_a network module, and features with different receptive fields are extracted through Atrous dilation convolution with different sampling steps in the RFB_a network module; the The extracted features of different receptive fields are input into the 3×3 convolutional layer to form the first scale prediction layer of the DSOD network; Step S3: After adding the Atrous convolutional layer with a preset sampling step size to the feature extraction sub-network in the DSOD network, input the feature map of the feature extraction sub-network described in step S2 into the feature extraction sub-network with a preset sampling step. In the long Atrous convolutional layer, it is used to increase the receptive field of the feature map; at the same time, the features generated by the Atrous convolutional layer are input into the multi-scale prediction layer in the DSOD network to form 5 scale prediction layers; Step S4: Input the features of the first scale prediction layer of the DSOD network described in the step S2 and the 5 scale prediction layers described in the step S3 into the multi-task loss function L adding the IOG penalty term; Step S5: set the learning rate through the preheating strategy, and use the gradient descent algorithm to optimize the weights of all network layers in the DSOD network; set the sample size to reduce the hardware equipment requirements for training the DSOD network; The multi-task loss function L added to the IOG penalty item described in step S4 is specifically: Step S41: Calculate the prediction frame g _{iou_max} with the largest area intersection ratio between the prediction frame p output by the DSOD network and all the real frames G of the preset sample, and the formula is as follows:

Among them, g represents the real box, G represents the set of all real boxes, p represents the prediction box, P represents the set of all prediction boxes, box _g represents the area of the real box, and box _p represents the area of the prediction box; Step S42: Remove the real frame with the largest area of the intersection and ratio in step S41, then calculate the largest IOG penalty item between the predicted frame and the remaining real frame, and use the largest IOG penalty item as the _Liog loss function. The calculation formula is as follows:

Step S43: Weighted fusion of the _Liog loss function, the localization loss function _Lloc and the classification loss function L _conf function to form the final multi-task loss function L, the formula is as follows:

Among them, N represents the number of detected positive samples, α represents the weight of L _loc of the localization loss; the localization loss function L _loc adopts the xmooth _L1 loss; the classification loss function L _conf adopts the information cross entropy to calculate; the localization loss function L _loc The calculation is as follows:

in,

Indication function, indicating that the ith default box matches the jth category of the real box, l represents the position coordinates of the predicted box, pos represents the default box of the positive sample, and N represents the number of positive samples; smooth _L1 is calculated as follows:

The classification loss _Lconf is calculated as follows:

Among them, c represents the confidence of each category, Neg represents the negative sample, p represents the category, 0 represents the category is the background,

Represents the probability that the i-th prediction box is of category p, calculated as follows:

2. a kind of method of improving DSOD network according to claim 1, is characterized in that: in step S1, image preprocessing is specifically: Step S11: Crop the input image: first randomly select the input image, that is, the length and height of the cropped image, and then randomly select a number among 0.1, 0.3, 0.7, and 0.9 as the Jaccard coefficient threshold, and calculate all the values in the original image through the Jaccard coefficient. The similarity of the ground-truth box and the cropped image; Step S12: Judging whether the Jaccard coefficient of the real frame or the cropped image is greater than the Jaccard threshold randomly selected in step S11; if at least one real frame and the Jaccard coefficient of the cropped image are greater than the selected Jaccard threshold, and the center coordinates of the real frame are If it falls in the cropped image, the cropped image meets the requirements, otherwise it returns to step S11; the Jaccard coefficient is calculated as follows:

Among them, N represents the number of real boxes in the image, box _i represents the area of the ith real box, box _cut represents the area of the cropped image, and operator ∩ represents the calculation of the overlap area; Step S13: Perform left and right mirroring processing on the cropped image according to a preset probability T, and adjust the resolution of the mirrored image to 300×300 to obtain a mirrored image; Step S14: De-average the mirrored image by using a de-average method to obtain a de-averaged image. 3. a kind of method of improving DSOD network according to claim 1, it is characterized in that: the concrete content of described step S2 is: at first use 1 × 1 convolution layer on each RFB_a network branch, in order to reduce the characteristic The number of channels; the first branch of the RFB_a network adopts a convolutional layer with a stride of 1 and a convolution kernel of 3 × 3 to obtain a 3 × 3 receptive field feature; the second branch adopts a 1 × 3 convolution layer and an Atrous convolutional layer with a sampling stride of 3 to obtain a 1×7 receptive field feature; on the third branch, a 3×1 convolutional layer and an Atrous convolutional layer with a sampling stride of 3 are used to obtain a 7×1 convolutional layer. 1 receptive field feature; adopt 3×3 convolutional layer and Atrous convolutional layer with sampling stride 5 on the fourth branch to obtain 11×11 receptive field feature; through channel stitching and 1×1 convolutional layer The features extracted by each branch are fused; finally, the fused features and the features generated by the second transition layer of the DSOD network are fused to form the final output features through residual fusion. 4. a kind of method of improving DSOD network according to claim 1, is characterized in that, adding described in step S3 has preset sampling step size and is the concrete method of Atrous convolution layer is: First, increase described DSOD network The number of output channels C of the feature extraction sub-network is used to extract richer feature information, and then the Atrous convolution layer with sampling step size r is added. The output channel of the Atrous convolution layer is equal to the original DSOD feature extraction sub-network. The number of output channels , so that Atrous convolution is embedded into the DSOD network; then a 1×1 convolution layer is added for feature fusion. 5. a kind of method for improving DSOD network according to claim 1, is characterized in that, described in step S5, setting learning rate by preheating strategy is specifically: setting initial learning rate to ^10-5 , in the first 5 The epoch increases the learning rate linearly to 10 ^-2 , divides the learning rate by 10 at the 75th epoch, 125th epoch and 175th epoch, respectively, and completes the training at the 200th epoch; batch normalized weight initial value It is set to 0.5, and the bias value is set to 0; all convolutions are initialized using the xavier method; the training batch sample size is reduced from 128 to 16 by improving the training strategy, in order to reduce the hardware equipment requirements for training the network.

Images