Fish Sonar Image Recognition Algorithm Based on Improved YOLOv5

Fish stock assessment is crucial for sustainable marine fisheries management in rangeland ecosystems. To address the challenges posed by overfishing of offshore fish species and to facilitate comprehensive deep-sea resource evaluation, this paper introduces an improved fish sonar image detection algorithm based on You Only Look Once version 5 (YOLOv5). Sonar image noise often results in blurred targets and indistinct features, thereby reducing the precision of object detection. Thus, the C3N module is designed in the neck component, where depth-separable convolution and an inverse bottleneck layer structure are integrated to lessen feature information loss during downsampling and forward propagation. Furthermore, A shallow feature layer is introduced in the network prediction layer to enhance feature extraction for pixels larger than 4x4. Additionally, Normalized Weighted Distance (NWD) based on a Gaussian distribution is combined with Intersection over Union (IOU) during gradient descent to improve small target detection and mitigate IOU's scale sensitivity. Finally, Traditional Non-Maximum Suppression (NMS) is replaced with Soft-NMS, reducing missed detections due to occlusion and overlapping fish targets common in sonar datasets. Experiments show the improved model surpasses the original model and YOLOv3 with gains in precision, recall, and mean average precision (mAP) of 2.3%, 4.7%, and 2.7%, respectively, and 2.5%, 6.3%, and 6.7%, respectively. These findings confirm the method's effectiveness in raising sonar image detection accuracy, consistent with model comparisons. With Unmanned Underwater Vehicles (UUV) advancements, this method holds the potential to support fish culture decision-making and facilitate fish stock resource assessment.