Lightweight single image super-resolution Transformer network with explicit global structural similarities capture

Lightweight networks for single image super-resolution (SISR) have increased in popularity, and numerous lightweight SISR networks have been proposed for various practical applications. However, most existing lightweight networks just explicitly capture local or regional range dependencies but cannot explicitly capture global range dependencies in an entire image due to the heavy computational costs. For this problem, we propose a lightweight patch Transformer network (PTN) for SISR, which explicitly captures the dependencies in local, regional and global ranges. Specifically, we propose a lightweight inter-patch Transformer (Inter-PT) layer to explicitly capture global structural similarities by summarizing structural information into a lower-dimensional space using the statistical properties of the average pooling operation. To better utilize local information, we design a deformable convolution module (DCM) to flexibly extract local features. Finally, a frequency reconstruction (FR) loss function is designed to recover more accurate high-frequency information in the reconstructed SR image. Extensive experimental results demonstrate that our proposed PTN can achieve better SR results than existing lightweight state-of-the-art SISR methods in terms of both performance and computational cost.