Fish sonar image recognition algorithm based on improved YOLOv5.

Area of Science:

• Marine Biology
• Computer Vision
• Fisheries Science

Background:

• Sustainable fisheries management relies on accurate fish stock assessment.
• Overfishing and challenges in deep-sea evaluation necessitate advanced detection methods.
• Sonar image noise degrades fish target features, reducing object detection precision.

Purpose of the Study:

• To introduce an improved fish sonar image detection algorithm based on YOLOv5.
• To enhance feature extraction and small target detection in noisy sonar images.
• To improve the accuracy and efficiency of fish stock assessment for marine fisheries.

Main Methods:

• Incorporated a C3N module with depth-separable convolution and inverse bottleneck layers into YOLOv5.
• Introduced a lowercase shallow feature layer for enhanced extraction of larger pixels.
• Combined normalized weighted distance with Intersection over Union (IoU) and replaced non-maximum suppression (NMS) with soft-NMS.

Main Results:

• The improved YOLOv5 model demonstrated significant gains in precision (2.3%), recall (4.7%), and mean average precision (2.7%) over the original model.
• Compared to YOLOv3, the enhanced model achieved higher improvements in precision (2.5%), recall (6.3%), and mean average precision (6.7%).
• The method effectively improved sonar image detection accuracy, particularly for small and overlapping targets.

Conclusions:

• The enhanced YOLOv5 algorithm significantly improves fish sonar image detection accuracy.
• This method offers a promising tool for fish stock assessment and marine resource evaluation.
• Advancements in Unmanned Underwater Vehicles can leverage this technology for better fisheries management and fish culture decision-making.