RSID: A Remote Sensing Image Dehazing Network

Hazy images often lead to problems such as loss of image details and dull colors, which significantly affects the information extraction of remote sensing images, so it is necessary to research image dehazing. In the field of remote sensing, remote sensing images are characterized by large-size and rich information, so the processing of remote sensing images often has the problems of GPU memory overflow and difficult removal of non-uniform haze. For remote sensing image characteristics, an efficient and lightweight end-to-end dehazing method is proposed in this paper. We use the FA attention combined with smoothed dilated convolution instead as the main structure of the encoder, which can achieve imbalanced handling of hazy images with different levels of opacity while reducing parameter count. Channel weight fusion self-attention is added in the decoder part to realize the automatic learning and pixel-level processing of different receptive field features We tested the proposed method on both public datasets RESIDE and real large-size hazy remote sensing images. The proposed method achieved satisfactory results in our experiments, which proves the effectiveness of the proposed method.