Single-Stage Underwater Target Detection Based on Feature Anchor Frame Double Optimization Network

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Underwater environments present significant imaging challenges, including color distortion, uneven lighting, blurring, and geometric distortion.
• These adverse conditions negatively impact the performance of object detection networks, leading to reduced accuracy and reliability.
• Standard object identification algorithms often fail to maintain robustness in underwater settings due to domain shift issues.

Purpose of the Study:

• To develop an improved single-stage object detection model specifically for challenging underwater environments.
• To enhance the resilience and accuracy of underwater object detection by optimizing feature representation and anchor box mechanisms.
• To address the limitations of existing methods in detecting small or occluded underwater objects.

Main Methods:

• A novel single-stage detection method, UWNet, is proposed, featuring dual enhancements to anchor boxes and feature representations.
• A composite-connected backbone network is utilized to improve feature context relevance and extraction capabilities.
• An receptive field enhancement module is incorporated to boost multi-scale detection performance, and a prediction refinement strategy refines anchor boxes and features through iterative regression.

Main Results:

• UWNet achieved a mean Average Precision (mAP) of 80.2% on the Labeled Fish in the Wild dataset, demonstrating improved accuracy.
• The model showed a 2.1 AP improvement over baseline methods, attributed to advanced feature extraction and multi-scale modules.
• At a 300x300 input resolution, UWNet attained 32.4 AP, with experimental results indicating that six prediction layers outperform four for this dataset.

Conclusions:

• The proposed UWNet model effectively enhances single-stage underwater object detection through optimized anchor boxes and features.
• The integration of three functional modules significantly boosts detection performance, particularly for small underwater targets.
• UWNet offers a robust solution to overcome common detection failures in complex underwater visual conditions.