Design of Forward-Looking Sonar System for Real-Time Image Segmentation With Light Multiscale Attention Net

Forward-looking sonar is a commonly used underwater detection device. However, due to the complex underwater environment, small target areas, and blurred features, the detection accuracy is poor and not conducive to quickly finding the desired targets. In recent years, deep learning has developed rapidly, and semantic segmentation network models have shown strong segmentation performance, with enormous application potential in forward-looking sonar image segmentation. A lightweight multiscale attention network is proposed in this article for real-time semantic segmentation of forward-looking sonar system. The proposed model ensures the training accuracy and inference speed of the semantic segmentation network through the structural reparameterization module and improves the feature extraction ability of the network by fusing multiscale information through the multiscale feature attention gate. Compared with other models, the proposed model achieved good results on semantic segmentation datasets and data collected from self-developed devices. The mean intersection over union (mIoU) reached 0.734 on a public dataset, outperforming existing mainstream models. Moreover, the inference speed can reach 1250 frame rate per second (FPS) on 2080Ti. An embedded forward-looking sonar host computer is designed based on Atlas200, and the proposed network can run up to 87 FPS on the embedded host computer, with a frame rate of 9.5 FPS per unit of power consumption, which is better than that of mobile phones and X86 computers. The forward-looking sonar system constructed based on this achieved satisfactory results in engineering practice.