KNN Local Attention for Image Restoration-Supplementary material

摘要: We first provide the performance comparison with stateof-the-art image restoration methods with respect to the accuracy and computational cost. Fig. 1 shows the graphs illustrating both the performance and computational cost of state-of-the-art methods. The proposed method is marked with a star symbol with red color, and other methods are marked with a circle symbol with green color. The x-axis and y-axis of the graphs represent the computational cost measured with Multiply-Accumulates (MACs) and the performance with the PSNR, respectively. The MACs of all graphs are measured when an input resolution is 256× 256. In the image denoising on the SIDD dataset [1], the proposed method has comparable computational cost with Uformer [14] and NBNet [5], while achieving the best performance. In the image deraining and deblurring, the KiT shows a slightly better performance yet with much less computational cost. Compared to the MPRNet [16], the KIT has almost 92.7% fewer MACs.