Improved Underwater Small Target Detection Algorithm for YOLOv7-Tiny

The appearance and shape of underwater targets are quite different from that of land targets, and traditional target detection algorithms are often difficult to accurately detect underwater targets and obtain more accurate information. In order to solve the above problems, this paper proposes an underwater small target detection algorithm CL-YOLOv7-tiny based on improved YOLOv7-tiny, which firstly replaces the original activation function by introducing a Sigmoid weighted linear unit (SiLU) activation function, adds a low-resolution self-attention module (LRSA) in Neck to capture the underwater key features of the target and integrate it into YOLOv7-tiny to further optimize the network structure and parameter settings to improve the accuracy and robustness of underwater target detection. The WIoU loss function is used to replace the loss function in the original YOLOv7-tiny network model, which improves the model generalization performance and facilitates the intelligent detection of underwater targets. Finally, Dyhead, a multi-head self-attentive detection head, is introduced at the output to improve the detection performance of small underwater targets at the output. Ablation experiments and comparison experiments are carried out on the underwater dataset, and the improved algorithm improves the mAP by 2.87 percentage points compared with the original Yolov7-tiny algorithm, which shows that the improved algorithm can effectively improve the efficiency of underwater small target detection to a certain extent.