Cross-Scale KNN Image Transformer for Image Restoration

Numerous image restoration approaches have been proposed based on attention mechanism, achieving superior performance to convolutional neural networks (CNNs) counterparts. However, they do not leverage the model in a form fully suited tasks. In this paper, we propose an network with novel called cross-scale k -NN Transformer (CS-KiT), that effectively considers several factors such as locality, non-locality, and aggregation, which are essential restoration. To achieve locality CS-KiT builds -nearest neighbor relation of local patches aggregates similar through attention. induce ensure each patch embraces different scale information scale-aware embedding (SPE) predicts input combination multi-scale convolution branches. We show effectiveness experimental results, outperforming state-of-the-art denoising, deblurring, deraining benchmarks.