MCRFS-Net: single image dehazing based on multi-scale contrastive regularization and frequency selection

• Published in: Scientific reports Vol. 15; no. 1; pp. 25501 - 14
• Main Authors: Qin, Qin, Shui, Lin, Zhang, Yanyan, Song, Shaojing, Jiang, Jinhua
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 15.07.2025; Nature Publishing Group; Nature Portfolio
• Subjects: 639/166/987; 704/106/35; Contrastive learning; Curricula; Deep learning; Fog; Humanities and Social Sciences; Image dehazing; Light; Multi-scale fusion mechanisms; multidisciplinary; Non-uniform haze; Preservation; Regularization methods; Science; Science (multidisciplinary); Image dehazing; Multi-scale fusion mechanisms; Contrastive learning; Non-uniform haze
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-025-08690-z

The primary goal of image dehazing is to restore the clarity and detail of hazy images. However, addressing non-uniform haze in atmospheric scattering models remains a significant challenge. While some methods tackle image-level non-uniformity using multi-scale fusion mechanisms, others focus on feature-level non-uniformity with content-guided attention mechanisms. However, few approaches effectively address non-uniform haze at both the image and feature levels simultaneously. To overcome this limitation, this paper introduces a novel Adaptive Multi-Scale Frequency Selection (AMFS) module, which consists of an Adaptive Multi-Scale Module (AMSM) and a Frequency Selection Block (FSB). The AMSM dynamically integrates multi-scale features through weighted fusion, effectively mitigating issues caused by non-uniform dehazing. Meanwhile, the FSB processes features in the frequency domain, highlighting critical high-frequency and low-frequency components via an attention mechanism, thereby enhancing detail preservation and suppressing noise. Additionally, we propose a Multi-Scale Contrast Regularization (MSCR) loss function, which leverages cross-scale contrastive learning to improve feature consistency. Experimental results demonstrate that the proposed algorithm outperforms existing methods on four benchmark datasets, achieving superior detail preservation and enhanced robustness against non-uniform haze.