Seafood Object Detection Method Based on Improved YOLOv5s.

Area of Science:

• Computer Vision
• Machine Learning
• Marine Biology

Background:

• Traditional underwater object detection algorithms struggle with false positives and missed detections.
• Accurate identification of aquatic species is crucial for marine research and resource management.

Purpose of the Study:

• To improve the accuracy and efficiency of underwater seafood object detection.
• To reduce false positives and missed detections in aquatic target identification.

Main Methods:

• Developed an improved detection method based on YOLOv5s.
• Introduced a Spatial-Channel Synergistic Attention (SCSA) module to enhance target features and suppress background noise.
• Integrated a three-scale convolution dual-path variable-kernel module (C3k2-PSConv) to improve multi-dimensional feature extraction, especially for small or occluded targets.

Main Results:

• The enhanced YOLOv5s model achieved a 2.3% increase in mean average precision (mAP) on the URPC dataset.
• Reduced the number of model parameters by approximately 2.4%, maintaining real-time inference speed.
• Demonstrated significant improvement in operational efficiency for underwater object detection.

Conclusions:

• The proposed method effectively enhances the detection of aquatic targets in complex underwater environments.
• The integration of SCSA and C3k2-PSConv modules offers a promising approach for real-time, accurate underwater object detection.
• The improved model shows potential for applications in marine monitoring and fisheries.