Unsupervised Image Dehazing Based on Improved Generative Adversarial Networks

Abstract

Image dehazing is a technique used for repairing blurry images which can effectively reduce the impact of haze on visual tasks. Most of the existing dehazing methods rely on atmospheric models or perform supervised learning based on paired images to obtain haze-free images. However, problems such as relying on prior knowledge of a specific scene and difficulty in collecting paired hazy and haze-free images have hindered the development of image dehazing techniques. In response to the above problems, we are inspired by the CycleGAN algorithm and propose the DAM-CCGAN algorithm, which uses an unsupervised method to dehaze unpaired images. For the blur and color distortion problems which can occur in image dehazing, the DAM-CCGAN algorithm adds a skip connection method and an attention mechanism module (DAM) to the generator. To preserve more image information, we add a detailed perception loss function. Meanwhile, to reduce the complexity of the algorithm, we improve the convolution group structure in the generator. Experiments show that our model achieves a good dehazing effect on both indoor and outdoor hazy images.