Positive-and-Negative Learning for Single Image Dehazing

Abstract

Single image dehazing has been a challenging problem for following high-level computer vision task due to the terrible information distortion. We noticed that most of existing dehazing methods are tend to pay more attention to positive sample and ignore crucial effect of negative sample. In this paper, we proposed a single image dehazing method that learns both positive sample (i.e. clean image) and negative sample (i.e. hazy image) to restore a hazy image, named Positive-and-negative Learning Dehazing Network (PNLDN). Firstly, we design a Positive Learning Dehazing Network by adopting Teacher Network based on knowledge distillation technology. Positive knowledge is absorbed by Dehazing Network. Secondly, to generate a better dehazing image, we proposed a Negative Contrast Optimization (NCO) module to push the dehazing output away from hazy image. Thirdly, we improved the common downsampling and upsampling method by using a deformable RoI (region of interest) pooling layer, named Dynamic Enhanced Sampling (DES). DES assists CNN to receive more spatial information of image and process irregular edge of objects. Experimental results have demonstrated that PNLDN surpasses state-of-the-art single image dehazing methods.