LW-YOLOv7SAR:Lightweight SAR Image Object Detection Method
Aiming at the characteristics of small targets,heavy noise,complexity in SAR scenarios,combined with the requirements of optimized lightweight detection models in ship target scenarios,we propose LW-YOLOv7SAR,a lightweight detection network for SAR ship image detection by pruning and optimization of the YOLOv7-tiny framework.It lightens the model through re-parameteriza-tion and Shuffle techniques,combined with GhostConv module and borrowing ideas and methods for removing redundant information.Meanwhile,it enhances the efficiency of multi-scale information extraction of the model.For ease of deployment and portability,the model uses easy-to-deploy activation functions such as hard-Swish and ReLU6.In addition,a soft thresholding module combined with spatial channel attention is introduced at the backbone layer to increase the denoising and generalization capabilities of the model.In order to improve the detection accuracy of small targets,a weighted multi-scale feature fusion module is introduced into the model.Through theoretical analysis and experimental verification,it is proved that the LW-YOLOv7SAR model reduces the calculation cost by 89％,the parameter amount by 90％,and the weight file size by 90％,compared with YOLOv7-tiny.Due to reducing the computa-tional demand,this model reduced power consumption during inference compared to the original model.Therefore it is also more in line with green computing requirements.The detection accuracy on the SSDD dataset can reach 97.6％.
NETL
NSTL
万方数据
