A Study on Remote Sensing Image Sharp-en-ing Using Graph Proxy Attention

Abstract

In recent years, as a technology that simulates the selective focusing ability of the human visual system, attention mechanisms have gained significant attention in the field of image processing. Within this field, Transformer architectures based on self-attention mechanisms show great po-tential for effectively addressing issues related to long-range dependencies and global feature modeling. Nevertheless, the high computational complexity associated with self-attention mechanisms has restricted their application. This paper introduces a novel attention mecha-nism—Graph Proxy Attention. It is applied in the remote sensing image sharpening model called HyperTransformer. Experiments are carried out on the public Pavia Center dataset. The im-proved model is evaluated in terms of computational efficiency and image sharpening perfor-mance. The results indicate that, compared to models based on self-attention and axial attention, the Graph Proxy Attention model significantly reduces the number of parameters and float-ing-point operations. Specifically, when compared to self-attention models, the parameter amount is reduced by 21.0%, and the floating-point operations are reduced by 22.6%. In com-parison to axial attention models, the parameter amount is reduced by 13.4%, and the float-ing-point operations are reduced by 15.7%. At the same time, the evaluation indicators for im-age sharpening quality have also achieved favorable results. For example, the relative global synthetic error has dropped to 0.48. These findings demonstrate the applicability of the Graph Proxy Attention mechanism in remote sensing image sharpening.