Synthetic aperture radar (SAR) image target detection method
The present disclosure provides a synthetic aperture radar (SAR) image target detection method. The present disclosure takes the anchor-free target detection algorithm YOLOX as the basic framework, reconstructs the backbone feature extraction network from the lightweight perspective, and replaces the depthwise separable convolution in MobilenetV2 with one ordinary convolution and one depthwise separable convolution. The number of channels in the feature map is reduced by half through the ordinary convolution, features input from the ordinary convolution are further extracted by the depthwise separable convolution, and the convolutional results from the two convolutions are spliced. The present disclosure highlights the unique strong scattering characteristic of the SAR target through the attention enhancement pyramid attention network (CSEMPAN) by integrating channels and spatial attention mechanisms. In view of the multiple scales and strong sparseness of the SAR target, the present disclosure uses an ESPHead.
1. A synthetic aperture radar (SAR) image target detection method, comprising the following steps:
performing feature extraction on an SAR image with a backbone feature extraction network;
inputting a feature extraction result to an attention enhancement pyramid attention network;
performing, by the attention enhancement pyramid attention network, parameter aggregation with features of different scales to extract multiscale information of a target object in the SAR image; and
extracting, by a target detection head, spatial hierarchical information of different reception fields from the multiscale information of the target object through different dilation rates.
2. The SAR image target detection method according to claim 1 , further comprising:
taking a lightweight network mobilenetV2 as a basic feature extraction network; and
constructing the backbone feature extraction network with an anchor-free target detection algorithm YOLOX based on the basic feature extraction network.
3. The SAR image target detection method according to claim 2 , wherein the constructing the backbone feature extraction network with the anchor-free target detection algorithm YOLOX based on the basic feature extraction network comprises:
replacing a depthwise separable convolution in the lightweight network MobilenetV2 with one ordinary convolution and one depthwise separable convolution;
reducing a number of channels in a feature map by half through the ordinary convolution, and extracting features of the feature map from the ordinary convolution through the depthwise separable convolution; and
acquiring the backbone feature extraction network.
4. The SAR image target detection method according to claim 1 , wherein the performing, by the attention enhancement pyramid attention network, parameter aggregation with features of different scales to extract multiscale information of the target object in the SAR image comprises:
selecting a squeeze-and-excitation module as a channel attention module in the attention enhancement pyramid attention network;
processing the feature extraction result through the channel attention module to perform differential weighting on different channels of a feature map; and
extracting the multiscale information of the target object in the SAR image.
5. The SAR image target detection method according to claim 4 , wherein the processing the feature extraction result through the channel attention module to perform differential weighting on different channels of the feature map comprises:
mapping the feature map in the feature extraction result to a size of 1*1*c, c being a number of channels;
learning weighting coefficients of the channels with two convolutional layers each having a kernel size of 1*1;
respectively multiplying the weighting coefficients of the channels by original features; and
convoluting, with a 1*1 convolutional layer, weighting coefficients of the channels obtained by multiplying the original features.
6. The SAR image target detection method according to claim 4 , wherein the extracting the multiscale information of the target object in the SAR image comprises:
splitting the feature extraction result into g groups through group convolution;
mapping, by split groups, the feature map to a size of 1*1 through Global Average Pooling, and seeking a dot product of each of grouped feature maps;
summating the dot product along channel-wise features to obtain a spatial attention mask which is as big as the original feature map and comprises one channel;
mapping, by the spatial attention mask, weights of spatial pixels to 0-1 through a sigmoid operation, and multiplying the weights by the original feature map;
re-splicing, by a concat operation, g groups of features to a size of the original feature map;
performing feature extraction on a splicing result with a convolution having a kernel size of 3*3; and
changing a convolutional reception field with convolutions having same kernel sizes and atrous rates of 1, 2, 4, 8, 16 to obtain the multiscale information of the target object.
7. The SAR image target detection method according to claim 1 , further comprising: predicting class information and position regression information of the target object in the SAR image with a Decoupled head.