A region-aware spatial-frequency adaptive fusion framework for robust single image dehazing

Atmospheric haze diminishes image visibility, contrast, and obscure fine details, adversely affecting visual perception and high-level vision tasks. Although prior-based dehazing methods have shown promising performance, these approaches relying solely on either the dark channel or bright channel prior estimates. These estimates, often fail to handle both non-sky and sky regions simultaneously. This paper proposes a region-aware spatial-frequency adaptive fusion framework. It integrates the complementary strengths of dark and bright channel priors transmission maps for effective scene radiance recovery. The fused image is further enhanced using Non-Subsampled Contourlet Transform (NSCT) based processing, incorporating gamma correction for low frequency components and wavelet-based denoising for high frequency components. Extensive experiments on five publicly available datasets demonstrate consistent performance based on PSNR, SSIM and other image-quality metrics. The proposed method demonstrates competitive performance with several conventional and deep learning-based dehazing approaches.