PESSANet: prior-enhanced selectable sparse attention network for image super-resolution
摘要
Transformer-based super-resolution (SR) methods have exhibited impressive performance in the field of image super resolution reconstruction. However, these methods face limitations in effectively utilizing prior information and involve a substantial number of irrelevant tokens during computation, which degrade the quality of image reconstruction. To address these issues, this paper proposes the prior-enhanced selectable sparse attention network (PESSANet), a novel framework specifically designed for SR tasks. The hybrid difference convolution block (HDCB) is employed for shallow feature extraction, enabling the capture of rich prior information, thereby enhancing the recovery of image textures. The local-global feature extraction block (LGEFB) is introduced to extract deep features, which consists of multi-head dynamic local self-attention block (MHDLSAB) and top-k selectable sparse attention block (KSSAB). Moreover, KSSAB adopts a more selectable and flexible strategy, which selectively retains the most relevant self-attention values, effectively reducing the impact of irrelevant tokens. The suggested method continuously surpasses current state-of-the-art techniques based on convolutional neural networks (CNNs) and Transformers, producing better results in both image quality and reconstruction accuracy, according to extensive experiments conducted on several benchmark datasets. Extensive experiments are conducted on several widely used benchmark datasets, including Set5, Set14, BSD100, Urban100, and Manga109, under multiple upsampling factors (e.g.,