An End-to-End Network for Single Image Dedusting

Dust storm not only reduces visibility of natural scene, but also affects capture capability of outdoor monitoring equipment. Image dedusting aims to remove dust storm in images and make image scenes clearer. However, most of existing image enhancement-based methods have focused on defogging and deraining, and have paid little attentions to image dedusting. Therefore, research of image dedusting is necessary. In this paper, we presented an end-to-end network for single image dedusting. First, the proposed method uses an encoder module to encode a dusty image into image feature maps. Second, it uses the latest smoothed dilated residual convolution module to alleviate gridding artifacts caused by a widely-used dilated convolution. Third, it adopts an attention-based module to fuse image features at different levels and extract more contextual image information. Finally, it utilizes a decoder module, which decodes the feature maps, to obtain a residue, and adds it to the input dusty image to obtain the final image dedusting result. Experiments demonstrate the effectiveness of the proposed method in image dedusting.