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

Single-image dehazing suffers from severe information loss and the under-constraint problem. The lack of high-quality robust priors leads to limited generalization ability of existing dehazing methods in real-world scenarios. To tackle this challenge, we propose a simple but effective single-image dehazing network by fusing high-quality semantic priors extracted from Segment Anything Model 2 (SAM2) with different types of advanced convolutions, abbreviated SAM2-Dehaze, which follows the U-Net architecture and consists of five stages. Specifically, we first employ the superior semantic perception and cross-domain generalization capabilities of SAM2 to generate accurate structural semantic masks. Then, a dual-branch Semantic Prior Fusion Block is designed to enable deep collaboration between the structural semantic masks and hazy image features at each stage of the U-Net. Furthermore, to avoid the drawbacks of feature redundancy and neglect of high-frequency information in traditional convolution, we have designed a novel parallel detail-enhanced and compression convolution that combines the advantages of standard convolution, difference convolution, and reconstruction convolution to replace the traditional convolution at each stage of the U-Net. Finally, a Semantic Alignment Block is incorporated into the post-processing phase to ensure semantic consistency and visual naturalness in the final dehazed result. Extensive quantitative and qualitative experiments demonstrate that SAM2-Dehaze outperforms existing dehazing methods on several synthetic and real-world foggy-image benchmarks, and exhibits excellent generalization ability.