Underwater Small Target Detection Based on Improved YOLOv7

An improved YOLOV7 target detection model is proposed for the problems of small target detection accuracy and missed detection in the process of underwater target detection. The simple parameter-free attention module (SimAM) is used to improve the focus of the network on small-scale targets. The lightweight operator content-aware reassembly of features (CAFAFE) is used to replace the nearest neighbor interpolation in the upsampling part to enhance the feature fusion capability of small targets. On these basis, the FocalEIOU loss function is used to replace the CIOU coordinate loss function in the original YOLOV7 network model to solve the sample imbalance problem, and reduce the miss detection rate. Making suitable datasets and conducting experiments, the results show that for the images in the dataset under three conditions of dark environment, small target and overlap, the improved YOLOV7 network model has a lower miss detection rate and significantly higher accuracy compared with the original network model. The mean average precision (mAP) has improved from 78.4% to 81.3%, compared with the original network model by 2.9 percentage points. The detection effect is improved over both the original network model and other traditional target detection network models.