Spatial Residual for Underwater Object Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Feature drift, arising from target features and degradation factors, diminishes underwater object detector performance.
• Existing methods struggle with the dynamic instability of target features under boundary constraints.

Purpose of the Study:

• To redefine feature drift within the context of partial differential equations (PDEs).
• To introduce a novel Spatial Residual (SR) block to mitigate feature drift and enhance detector performance.

Main Methods:

• Proposed the Spatial Residual (SR) block, utilizing SkipCut for network-wide constraints and optimization.
• Implemented the SR block as a 5-layer backbone (BSR5) for complex feature extraction.
• Integrated BSR5 into DETR and YOLO architectures for performance evaluation.

Main Results:

• BSR5-integrated DETRs and YOLOs achieved state-of-the-art results on the RUOD dataset.
• BSR5-DETR demonstrated improved AP (1.3% and 2.7%) over RT-DETR with ResNet-101, while reducing parameters.
• BSR5 showed strong convergence and robustness, particularly in training from scratch.

Conclusions:

• The proposed SR block effectively addresses feature drift in underwater object detection.
• BSR5 offers a general-purpose backbone suitable for data-scarce, resource-constrained, and real-time applications.
• The SkipCut mechanism optimizes information flow and gradient allocation, accelerating training and enhancing performance.