Underwater Object Detection Using TC-YOLO with Attention Mechanisms.

Area of Science:

• Computer Vision
• Robotics
• Marine Technology

Background:

• Underwater object detection is crucial for intelligent underwater vehicles.
• Challenges include blurry images, small/dense targets, and limited onboard computation.
• Existing methods struggle with these specific underwater conditions.

Purpose of the Study:

• To enhance underwater object detection performance for intelligent vehicles.
• To address challenges like image blurriness, target density, and computational constraints.
• To develop a robust and efficient detection model suitable for mobile underwater platforms.

Main Methods:

• Proposed a novel detection network, TC-YOLO, based on YOLOv5s.
• Integrated Transformer self-attention and coordinate attention for improved feature extraction.
• Employed adaptive histogram equalization for image enhancement.
• Utilized optimal transport for effective label assignment.

Main Results:

• TC-YOLO demonstrated superior performance over YOLOv5s and other networks on the RUIE2020 dataset.
• The model effectively handles blurry images and dense targets.
• Ablation studies confirmed the contribution of each proposed component.
• The model maintains a small size and low computational cost.

Conclusions:

• The proposed TC-YOLO approach significantly improves underwater object detection.
• The combination of network architecture, image enhancement, and label assignment is effective.
• The model is suitable for real-time applications on resource-constrained underwater vehicles.