Real-time ship detection system for wave glider based on YOLOv5s-lite-CBAM model

The on-line ship detection system based on wave glider has good application prospect. However, there are still challenges in model weight, recognition accuracy and real-time performance when the system is applied in the remote ocean environment. Hence, a real-time ship detection system based on YOLOv5s-lite-CBAM model is proposed in this paper. Additionally, a JPEG-PNG image compression algorithm is introduced to compress the satellite return pictures effectively. The C3 in YOLOv5s is replaced by the combination module of shuffleNetV2 and CBAM's attention mechanism, which greatly reduce the weight of the model improve the detection accuracy. To optimize image compression, JPEG and PNG algorithms are combined to preserve the ship information even at high compression rates. The training results demonstrate that the proposed YOLOv5s-Lite-CBAM model achieves a 2.1 % increase in mAP and reduces the weight file from 13.6 MB to 3.5 MB compared to the conventional YOLOv5s. The system has been installed on a wave glider and operated steadily over the remote ocean for more than four months. The sea trials under smooth sea states show that the YOLOv5s-Lite-CBAM model exhibits a 10.4 % improvement in detection capability and a 2.9 % reduction in false detection rate compared to YOLOv5s.