A guiding teaching and dual adversarial learning framework for a single image dehazing

In most existing deep learning-based image dehazing methods, the haze-free source images are only used as the ground truth for the design of the loss function, whereas the guiding role that the source image should play on different feature levels has been ignored. This will result in a sub-optimal dehazing output. To address this issue, inspired by the knowledge distillation, a guiding teaching framework is designed for single image dehazing in an end-to-end manner, where the features of the haze-free source image at different levels are completely used to promoting the restoration of the hazy image. Specifically, the framework consists of a two-stream convolutional neural network termed teacher stream (TS) and student stream (SS), respectively. The input of the former is a haze-free image while the output is the desired image after reconstruction. The input of the latter is the hazy image, and the output is the restored image. Moreover, a dual adversarial strategy is designed to further improve the ability of SS to imitate teacher stream. In this process, the output results of SS are divided into two categories according to their hazy intensity levels. Then a thick light discriminator is introduced and made against the SS pit, such that the images with better dehazing effects can be used to deal with the ones poorly dehazed. A second discriminator termed light clear discriminator (LCD) is further introduced and a minimax game between the LCD and the SS is defined to drive the final result produced by SS closer to the reconstruction result of the TS. Experimental results show that the proposed method outperforms several latest methods applied to both artificial hazy images and the hazy images from the real scene.