A Lightweight Convolutional Neural Network for Ship Target Detection in SAR Images

Synthetic aperture radar (SAR), due to its merits of all-day, all-weather, and high resolution, has become an important component of radar research. Target detection is a significant basis for radar image interpretation. Aiming at the problems, such as low accuracy and high complexity of traditional convolutional neural network in SAR ship target detection, which is not conducive to the deployment of terminal equipment, we proposed a lightweight network for SAR image ship target detection based on the latest anchor-free algorithm YOLOX. First, we redesign a novel lightweight block with stronger feature fusion ability, namely, MobileNetV3S. On the basis of MobileNetV3S, we combine Cross Stage Partial Network to construct a lightweight backbone. Second, in order to improve the ability of multiscale feature extraction, we propose a new module based on dilated convolution with different dilated rates and Efficient Spatial Pyramid Network. Furthermore, we adopt convolutional block attentional module to optimize traditional YOLOX's Feature Pyramid Network, and propose a lightweight enhanced feature extraction module, which can improve the focusing ability of important targets. For detection head, the depth-separable convolution is also applied to reduce the network's parameters. Finally, in terms of loss function, we abandon the traditional Intersection over Union and use absolute Intersection over Union with the better convergence effect. The experimental results on the SAR Ship Detection Dataset show that compared with the baseline YOLOX, although parameters of our method are decreased by 66.7%, its AP reaches 90.8%, which exceeds the baseline YOLOX by 0.5%, and its false detection rate is also obviously reduced, achieving state-of-the-art performance.