SAM2-Dehaze: Fusing High-Quality Semantic Priors with Convolutions for Single-Image Dehazing.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Single-image dehazing faces challenges like information loss and under-constrained problems.
• Existing methods struggle with generalization due to a lack of robust priors.
• Real-world foggy images require advanced dehazing techniques for clarity.

Purpose of the Study:

• To propose a simple yet effective single-image dehazing network, SAM2-Dehaze.
• To leverage high-quality semantic priors from Segment Anything Model 2 (SAM2) for improved dehazing.
• To enhance the generalization ability of dehazing methods in real-world scenarios.

Main Methods:

• Developed SAM2-Dehaze, a U-Net-based network incorporating advanced convolutions.
• Utilized SAM2 for generating accurate structural semantic masks.
• Designed a dual-branch Semantic Prior Fusion Block for feature collaboration.
• Introduced a novel parallel detail-enhanced and compression convolution.
• Incorporated a Semantic Alignment Block for post-processing.

Main Results:

• SAM2-Dehaze demonstrates superior performance on synthetic and real-world foggy benchmarks.
• The network achieves excellent generalization ability across diverse foggy image datasets.
• Quantitative and qualitative experiments validate the effectiveness of the proposed method.

Conclusions:

• SAM2-Dehaze effectively addresses information loss and under-constraint issues in single-image dehazing.
• The fusion of SAM2 semantic priors and advanced convolutions significantly enhances dehazing performance.
• The proposed method offers a robust and generalizable solution for real-world foggy image restoration.