Sparsity regularization-based real-time target recognition for side scan sonar with embedded GPU

Abstract

Side Scan Sonar (SSS) is widely used to search for seabed objects such as ships and wrecked aircraft due to its high-imaging-resolution and large planar scans. SSS requires an automatic real-time target recognition system to enhance search and rescue efficiency. In this paper, a novel target recognition method for SSS images in varied underwater environment, you look only once (YOLO)-slimming, based on convolutional a neural network (CNN) is proposed. The method introduces efficient feature encoders that strengthen the representation of feature maps. Channel-level sparsity regularization in model training is performed to speed up the inference performance. To overcome the scarcity of SSS images, a sonar image simulation method is proposed based on deep style transfer (ST). The performance on the SSS image dataset shows that it can reduce calculations and improves the inference speed with a mean average precision (mAP) of 95.3 and at least 45 frames per second (FPS) on an embedded Graphics Processing Unit (GPU). This proves its feasibility in practical application and has the potential to formulate an image-based real-time underwater target recognition system.