SMD-Hazeless++: Advancing Vision Mamba With Omni-Directional Scans and Gated Dynamic Convolutions for Holistic Image Dehazing
Mukhtar Ahmad, Arslan Shaukat, Muhammad Usman Akram, Muhammad Waqas Ahmad · IEEE Access · 2026
Single-image dehazing is inherently ill-posed and complex because haze varies spatially and must be removed while simultaneously maintaining fine details and restoring global visibility. To mitigate this ill-posed problem, an innovative image dehazing framework called SMD-Hazeless++ is proposed in this study. It extends our original Shallow-Medium-Deep level Hazeless Network(SMD-Hazeless) architecture by employing the hierarchical multi-level feature transformations and simulating the explicit haze phenomenon. At the shallow level, focus is to enhances fine-grained textures using dynamic directional and multi-scale convolutions (DDMSC). Subsequently, at medium level high and low frequency components are separated through Discrete Wavelet Transform (DWT) module. Within the Medium Level Feature Transformation (MLFT), the low-frequency components are processed by the Haze Density Estimation Map (HDEM) module to estimate local degradation distributions. Concurrently, the corresponding high-frequency sub-bands are passed through a localized DDMSC block to explicitly refine and restore structural details degraded by haz. Afterwards, it feeds the low frequency augmented with estimated map to deep level transformation which comprises of Haze-Aware Directional Mamba (HADM) module. This module effectively suppresses non-homogeneous haze degradation while maintaining long-range contextual dependencies and global scene information. Ultimately decoder integrates the three-tier feature transformations through attention-guided gated fusion and produces detail rich and aesthetically good dehazed image. The proposed model is trained using composite loss function to ensure both structural and visual fidelity. Extensive experiments on I-HAZE, O-HAZE, Dense-HAZE, NH-HAZE, RTTS, and URHI datasets demonstrate that SMD-Hazeless++ consistently outperforms contemporary methods on reference-dependent (PSNR, SSIM) and reference-free (MUSIQ, PI, MANIQA, and Q-Align) IQA metrics, achieving robust dehazing under both uniform and non-uniform haze conditions.