One stage multi-scale efficient network for underwater target detection.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Underwater target detection is challenging due to environmental complexity, leading to low precision for small or dense targets.
• Existing methods struggle with accuracy in complex underwater scenarios, necessitating advancements in detection algorithms.

Purpose of the Study:

• To propose a novel underwater target detection method based on YOLOv5s to enhance precision and robustness.
• To improve feature extraction, representation, and fusion for more accurate underwater object identification.

Main Methods:

• An efficient feature extraction network and a novel attention mechanism with deformable convolution were employed.
• An adaptive spatial fusion operation was introduced in the YOLOv5s neck for effective multi-layer feature fusion.
• An adaptive fusion feature pyramid network was utilized to integrate global semantic information and reduce the feature semantic gap.

Main Results:

• The proposed method achieved an mAP50 of 86.97% on the Underwater Robot Professional Contest of China 2020 dataset, a 3.07% improvement over YOLOv5s.
• The method demonstrated a detection precision of 76.0% on the PASCAL VOC2007 dataset, outperforming several existing methods.
• Experimental results confirm enhanced precision and robustness in underwater target detection.

Conclusions:

• The novel YOLOv5s-based method significantly improves underwater target detection accuracy, particularly for small and dense objects.
• The integration of advanced feature extraction, attention mechanisms, and adaptive fusion enhances detection performance in complex environments.
• The proposed approach offers a robust solution for underwater target detection, outperforming standard methods on benchmark datasets.