The Generative Adversarial Network Based on Attention Mechanism for Image Defogging

Abstract

AbstractThe image collected under the haze weather conditions has some obvious problems such as reduced contrast, reduced clarity and color distortion, which seriously reduces the image quality. Combining with deep learning, this paper proposes a new end-to-end generative adversarial defogging network based on attention mechanism, which transforms the problem of image defogging into the generation of fog-free images based on foggy images. On the basis of generative adversarial network, the attention mechanism was introduced into the generating network to enhance the features of the foggy region and generate the attention map. Then the attention map and the foggy image were input into the auto-encoder for encoding and decoding the image and re-decode the foggy. Finally, the fogless image was output. The attention map was introduced into the discriminate network to discriminate the true and false of the fogless image (the prediction fogless image and the real fogless image). A lot of experiments were carried out on the foggy images in different scenes, and the various comparative analysis were carried out with the evaluation index of image quality. The result shows that the defogging network model proposed in this paper can not only be better applied to all kinds of foggy scenes, but also generate fog-free images that are more suitable for real scenes, and effectively improve the average gradient value, information entropy and NRSS of the image, and can better restore edge information and color information of foggy images.KeywordsImage defoggingGenerative adversarial networkAttention in neural networkStructural similarity loss functionInduction loss function