Color Channel-Based Smoke Removal Algorithm Using Machine Learning for Static Images

Wei-Ting Chen, Shih-Yi Yuan, Gui-Cheng Tsai, Hui-Chih Wang, Sy-Yen Kuo

Abstract: Images acquired from digital cameras are usually interfered by smoke, which may degrade the performance of object detection. There are few algorithms focused on smoke removal for still images so far and we usually use haze removal algorithms to remove smoke instead. However, there exist some differences between haze and smoke (e.g. particle properties and localization). Thus, a dehaze algorithm usually has limited performance for smoke removal. In this paper, we propose a novel smoke removal algorithm based on machine learning and smoke detection techniques. Moreover, we observed that the intensity distributions are not the same for different color channels in smoky images. Therefore, the proposed algorithm trains the models corresponding to each color channel and remove smoke from RGB channels separately. Simulations show that the proposed algorithm can significantly remove smoke. Moreover, as far as we know, the proposed algorithm is the first smoke removal algorithm for static images.