Cross Scale Attention Transformer for Single Image Super-Resolution

Vision Transformer has demonstrated remarkable performance in single image super-resolution tasks, yet their reliance on uniform dimensionality for query, key and value limits their ability to capture multi-scale features efficiently. Existing methods reduce computational complexity or facilitate information interaction by selectively altering the dimensions of the three. For instance, Swin Transformer reduces computational complexity through window-based attention and Shunted Transformer adjusts different key and value dimensions to enhance feature interaction. However, these approaches fail to efficiently utilize interactive information across varying scales due to rigid dimensional constraints. In this work, we propose a novel Cross Scale Attention Transformer (CSAT), which dynamically modulates the length of query, key and value tensors to integrate hierarchical contextual cues. By decoupling and reconfiguring the dimensional configurations of attention matrics, our method enables fine-grained cross scale feature fusion while maintaining computational efficiency. Experiments show that our CSAT achieves outstanding performance compared to other methods.