A Lightweight Underwater Target Detection Network for Seafood

With the development of underwater vehicle technology, the application of underwater target detection ranges from ocean search, rescue and detection to fishing for marine products. At this stage, the main problem for seafood detection is the slow detection processing time caused by a large number of parameter calculations in deep learning networks. This paper proposes a lightweight underwater target detection network based on YOLOv5s. First, GhostBottleneck is used to reduce the network parameters. Second, a convolution block attention mechanism is added to the last layer of the backbone network to increase the weight of the valuable features while suppressing invalid features to improve detection accuracy. Finally, GIoU_ NMS in YOLOv5s replaces DIoU_NMS, which enhances the detection accuracy of the network for clustered overlapping targets. The results show that the lightweight network parameters and floating point calculation are 36.11% and 43.63% less than those of YOLOv5s, the FPS is 109.12, and the detection accuracy is increased by 1.53%. Fast and efficient detection considerably improves the efficiency of underwater seafood detection.