Efficient underwater object detection based on feature enhancement and attention detection head.

Area of Science:

• Computer Vision
• Marine Technology
• Artificial Intelligence

Background:

• Underwater object detection is crucial for ocean exploration and conservation.
• Existing algorithms struggle with poor illumination, scattering, and absorption in underwater images.
• Suboptimal detection performance hinders effective analysis of marine environments.

Purpose of the Study:

• To enhance the feature extraction and integration capabilities for underwater targets.
• To improve the localization accuracy of foreground objects in complex underwater scenes.
• To develop robust solutions for real-time underwater object detection.

Main Methods:

• Proposed a Partial Semantic Encoding Module (PSEM) to improve multi-scale feature fusion and local channel information encoding.
• Introduced a Split Dimension Weighting Head (SDWH) for enhanced spatial and channel semantic weighting.
• Integrated PSEM and SDWH into YOLO series algorithms (YOLOv5, YOLOv8, YOLOv6).

Main Results:

• YOLOv8n with PSEM and SDWH achieved a 2.8% mAP increase on the UTDAC2020 dataset.
• Significant mAP improvements were observed across YOLOv5n (1% to 1.5%) and YOLOv6n (3% to 3.7%) on both UTDAC2020 and RUOD datasets.
• Achieved state-of-the-art performance with YOLOv8n reaching 82.9% mAP on UTDAC2020 and 80.9% on RUOD.

Conclusions:

• PSEM and SDWH modules substantially enhance underwater object detection accuracy in YOLO detectors.
• The proposed methods offer significant performance gains with acceptable computational costs.
• The real-time capabilities of the enhanced models meet practical requirements for marine applications.