Enhanced YOLOv7 for Improved Underwater Target Detection

Abstract

Aiming at the problems of the underwater existence of some targets with relatively small size, low contrast, and a lot of surrounding interference information, which lead to a high leakage rate and low recognition accuracy, a new improved YOLOv7 underwater target detection algorithm is proposed. First, the original YOLOv7 anchor frame information is updated by the K-Means algorithm to generate anchor frame sizes and ratios suitable for the underwater target dataset; second, we use the PConv (Partial Convolution) module instead of part of the standard convolution in the multi-scale feature fusion module to reduce the amount of computation and number of parameters, thus improving the detection speed; then, the existing CIou loss function is improved with the ShapeIou_NWD loss function, and the new loss function allows the model to learn more feature information during the training process; finall, we introduce the SimAM attention mechanism after the multi-scale feature fusion module to increase attention to the small feature information, which improves the detection accuracy. This method achieves an average accuracy of 85.7% on the marine organisms dataset, and the detection speed reaches 122.9 frames/s, which reduces the number of parameters by 21% and the amount of computation by 26% compared with the original YOLOv7 algorithm. The experimental results show that the improved algorithm has a great improvement in detection speed and accuracy.