Research on Deep Learning Models for Dehazing Wildfire Images

In recent years, image dehazing tasks have primarily focused on urban or indoor scenarios, with little attention paid to the study of image dehazing in forest fire contexts. The degradation of images caused by the dense smoke generated by forest fires poses a significant obstacle to firefighting and rescue operations. Unlike ordinary haze, the smoke produced by forest fires is characterized by its non-homogeneous density, which complicates the process of smoke removal. Moreover, there is limited research dedicated to smoke removal specifically in wildfire environment images. To address this gap in image dehazing research for forest fire scenarios, this study trains and tests various specialized dehazing methods on a new public dataset for forest fire dehazing, Fo-haze. The methods are divided into four groups based on their network structures and training strategies. Through qualitative and quantitative evaluations of the models' performance, the study aims to assist firefighters in assessing the dehazing capabilities of different models. The experimental results indicate that the CHaIR model from the Technical University of Munich, Germany, achieved the best results, with a PSNR of 27.11 dB and an SSIM of 0.8721.