Single image dehazing via cycle-consistent adversarial networks with a multi-scale hybrid encoder-decoder and global correlation loss

Abstract

Deep learning technologies have been reshaping the research of image dehazing in recent years with superior performance. Due to the difficulty to obtain the real hazy and corresponding clear image pairs in the wild, most of the existed deep learning based methods utilize synthetic datasets for model training. As a result, the performance and robustness of those methods are compromised under the real-world complex scenarios. In this paper, we propose a novel image dehazing method for unpaired data via cycle-consistent adversarial networks with a multi-scale hybrid encoder-decoder and global correlation loss. The requirement of paired training data is eliminated by combining two generators and discriminators into a cycle-consistent adversarial network. Moreover, to further improve the feature representation capability of the network for degraded images, a multi-scale hybrid encoder-decoder structure is introduced into the generators and multiple residual and dense blocks are constructed. Furthermore, to preserve more details of color and structure in generated dehazed images, a global correlation loss function is proposed. The task-specific haze-line prior is reformulated in the form of color loss constraint and incorporates with adversarial loss, cycle consistency loss, identity mapping loss, and perceptual consistency loss into a unified framework. Comprehensive qualitative and quantitative experiments on both synthetic and real-world datasets demonstrate the favourable dehazing results of the proposed method compared with a number of state-of-the-art methods. The code will be made available on Github.