Image super-resolution algorithm based on multi-dimensional Swin Transformers

Numerous studies have demonstrated the Swin Transformer performs well in image super-resolution tasks. However, Swin Transformer typically divides the input image into fixed size blocks (such as 48x48) and independently processes each block, causing it to overlook a wider area. Furthermore, many deep learning super-resolution technologies face the challenge of computationally intensive operations. To address these issues, we propose a new NSTSR model consisting of four parts: data preprocessing layer, primary feature extraction layer, advanced feature extraction layer, and image generation layer. The advanced feature extraction layer consists of multiple STNBs (Swin Transformer N-segment blocks). We have introduced a technique called N -Segments, which applies the N -Gram contextual method for natural language processing tasks to low-level visual tasks. In our work, N-Segments refers to neighboring local windows in Swin, distinct from the traditional text analysis definition where N -Segments represents a continuous picture segment. Leveraging sliding-WSA, the N-Segments interact with each other, enabling the expansion of the regions seen to restore degraded pixels effectively. The experimental results demonstrate that the NSTSR model exhibits excellent performance compared to existing advanced methods while maintaining the computational efficiency structure.