Progressive Multi-Scale Perception Network for Non-Uniformly Blurred Underwater Image Restoration

Area of Science:

• Computer Vision
• Image Processing
• Robotics

Background:

• Underwater imaging suffers from spatially varying blur due to turbulence, scattering, and motion.
• Existing enhancement methods inadequately address non-uniform blur in realistic underwater scenarios.
• This limitation hinders performance in critical downstream vision tasks.

Purpose of the Study:

• To develop an advanced network for effectively handling non-uniform underwater blur.
• To improve visual quality and performance of underwater vision systems.
• To provide a benchmark for evaluating non-uniform underwater blur restoration.

Main Methods:

• Proposed PMSPNet (Progressive Multi-Scale Perception Network) for non-uniform underwater blur.
• Integrated Hybrid Interaction Attention Module for feature ambiguity modeling.
• Employed Progressive Motion-Aware Perception Branch and Progressive Feature Feedback Block for refined blur localization and reconstruction.
• Constructed a diverse Non-uniform Underwater Blur Benchmark.

Main Results:

• PMSPNet achieved superior performance compared to state-of-the-art methods on multiple datasets.
• Demonstrated an average PSNR of 25.51 dB.
• Achieved a fast inference speed of 0.01 seconds.

Conclusions:

• PMSPNet effectively addresses the challenge of non-uniform underwater blur.
• The method enhances visual perception for underwater robots, ecological monitoring, and inspection.
• PMSPNet offers a significant advancement in underwater image restoration technology.