<p>High-temperature steam and single-point lighting in fermentation tanks cause severe image degradation, characterized by dense haze and uneven illumination with local overexposure. These conditions violate the uniform illumination assumption inherent in most physical priors. Consequently, traditional algorithms, such as the Dark Channel Prior (DCP), often misinterpret bright reflections as haze, leading to inaccurate transmission estimation and color distortion. Although deep learning-based approaches have shown promise, they are severely limited by their dependence on large-scale paired training data. Collecting pixel-aligned clear reference images in such dynamic industrial environments is practically infeasible, rendering supervised methods inapplicable. To address these challenges, this paper proposes an image dehazing algorithm tailored for fermentation tanks. The method first segments the image into high- and low-illumination regions based on luminance distribution. Specifically, for high-illumination regions, a Structural-Information and Brightness Prior (SIBP) is proposed for robust atmospheric light estimation, which combines structural cues from guided filtering and quadtree-based brightness statistics. To eliminate boundary artifacts between regions, a Saliency- and Direction-Aware Wavelet Fusion (SDAWF) strategy is developed, which integrates saliency-guided low-frequency fusion with direction-aware high-frequency weighting. Comparative experiments demonstrate that the proposed method significantly enhances both structural fidelity and perceptual naturalness under complex illumination conditions.</p>