EHA-Transformer: Efficient and Haze-Adaptive Transformer for Single Image Dehazing

Deep learning based dehazing structures have achieved significant progress in image haze removal. However, most recent methods mainly focused on the excellent feature extraction and representation capabilities of deep networks, and neglected the contributions of traditional haze-relevant priors to image dehazing. In this paper, we propose a novel dehazing method, named EHA-Transformer, which fully integrates the Transformer with haze-relevant features and enhances the interpretability. Since the haze distributions vary in different regions, the difficulties of local patch dehazing are also different. Based on this, we first propose a haze detector to distinguish regions, which are prone to produce residual haze during dehazing. Then, we introduce a haze-adaptive loss into our dehazing framework to increase the stability of the training process. Our dehazing framework is simple and generic, and can be easily applied to current dehazing models without introducing complexity. Since our EHA-Transformer takes full account of haze related properties, comprehensive experiments compared with state-of-the-arts demonstrate our framework have significant improvements in terms of robustness. We also apply our framework into different backbones, the noticeable improvements of different dehazing backbones illustrate the generalization capability of our framework.