Image Super-Resolution Using Dilated Window Transformer

Transformer-based networks using attention mechanisms have shown promising results in low-level vision tasks, such as image super-resolution (SR). Specifically, recent studies that utilize window-based self-attention exhibited notable advancements SR. However, self-attention, a slower expansion of the receptive field, thereby restricting modeling long-range dependencies. To address this issue, we introduce novel dilated window transformer, namely DWT, which utilizes dilation strategy. We employ simple yet efficient strategy enlarges by inserting intervals between tokens each to enable rapid and effective field. In particular, adjust interval become wider layers go deeper. This enables extraction local features allowing interaction neighboring shallow while also facilitating global enabling not only adjacent but distant deep layers. conduct extensive experiments on five benchmark datasets demonstrate superior performance our proposed method. Our DWT surpasses state-of-the-art network similar sizes PSNR margin 0.11dB 0.27dB Urban100 dataset. Moreover, even when compared with about 1.4 times more parameters, achieves competitive for both quantitative visual comparisons.