Abstract: Image defogging is a challenging and hot issue in the field of computer vision. Existing defogging methods usually use a single convolutional neural network (CNN) to solve the problem, but such methods lack  detail recovery mechanism and perform poorly in the case of non-uniform fog. To address the above two problems, a dual-prior guided attention feature aggregation defogging generative adversarial network is proposed, where the dark channel prior and semantic prior respectively guide the recovery of generalized features and texture details of the images. The generator uses a parameter-sharing encoder to extract features, adds an attention feature aggregation block (AFAB) to aggregate and enhance multi-scale features, and recovers the fog-free image by decoding multi-scale features. Finally, a multi-scale discriminator is used to supervise the recovery of the fog-free image. In addition, considering the possible uneven distribution of fog in the image, a coordinate attention residual block (CARB) is proposed, which can adaptively assign weights to make the network focus on the important features of the image. At the same time, a coordinate attention dense residual group (CARG) is constructed through three CARBs using residual aggregation, so that residual features can be fully utilized.Experimental results show that the proposed network performs excellently on both synthetic foggy image datasets and real foggy image datasets.

Key words: image defogging, generative adversarial network（GAN）, dual-prior guided, attention feature aggregation, parameter-sharing encoder, coordinate attention