Towards real-time detection of underwater target with pruning lightweight deep learning method in side-scan sonar images

Real-time detection of underwater targets through side-scan sonar (SSS) images is challenging due to the complex underwater environment and heavy computational load. This paper proposes a pruning lightweight deep learning network for real-time underwater target detection. A feature extraction backbone network combining the receptive field convolution and the omni-dimensional dynamic convolution is designed to extract the diversified features of sparse and unclear targets in SSS images effectively. Furthermore, a lightweight feature information aggregation network is developed. It consists of a lightweight feature aggregation architecture, a feature information filtering and capturing module, and an upsampling module. On this basis, reliable pruning and knowledge distillation algorithms are introduced to reduce further the computational burden and parameter complexity for real-time detection. Experimental studies are conducted on the shipwreck SSS images dataset and public acoustic images dataset, which show that the developed lightweight network achieves accurate and stable detection of underwater targets. On the shipwreck SSS images dataset, the average accuracy mAP50 of underwater target detection reached 87.6% by applying the proposed model, whereas the baseline model is 83.9%. The number of parameters of the network is reduced by 56.65% compared with the baseline model, and the floating-point operations (FLOPs) are reduced by 68.29%. In addition, the dataset includes 851 SSS images of shipwrecks is created and available at https://gitee.com/nwpu-r/underwater-shipwreck-dataset.