Real-Time Dehazing Method in Industrial Scenes Based on Affine Bilateral Learning
Liang Huan Wu, Ronghui Liang, Juan Du, Tao Liang, Gang Yi, Xiaoyi Zhang · 2025
Many industrial scenes are affected by haze, which results in significant visual degradation. A typical example is the Refuse-Derived Fuel (RDF) bunker, where fermentation and disinfection of the waste cause large amounts of dust to remain suspended in the air, resulting in camera fogging and impairing various advanced visual tasks. Although there are various haze removal technologies available at present, most of these haze removal methods are developed based on natural haze models, so they generally have difficulty handling the non-uniform and dynamic haze commonly found in industrial scenarios. To solve this problem, we propose a real-time haze removal model, which roughly consists of three modules. Module 1 captures the hazy features of the input image and constructs a bilateral mesh. Module 2 reconstructs clear haze-removed features by combining different color channels with the bilateral mesh. Finally, module 3 fuses these features using channel and spatial attention to generate a clear output. Experimental results show that the proposed model achieves the SOTA performance in the RDF bunker scenario, and can achieve 34 frames per second when processing 720p images on a single NVIDIA GeForce RTX 4070 Ti GPU. Compared with the 4KDehaze baseline model, the proposed method reduces the BRISQUE and NIQE scores by 6.2% and 5.5% respectively, and increases the CEIQ by 3.0%.