High-rank corrected multi-head self attention for image super resolution

Recently, Transformer-based methods have shown impressive performance in image super resolution (SR) tasks, by exploiting multi-head self attention (MSA) to capture long-range dependencies between pixels. Unfortunately, there is a low-rank bottleneck in existing Transformer-based SR methods, which limits SR performance. We demonstrate that this is because the attention map in MSA is restricted to using more non-zero singular values to make stable representation. Increasing the projection dimension of MSA can eliminate the low-rank bottleneck, but results in overwhelming computational burden. Furthermore, we observe that the attention maps of different heads in MSA exhibit both information redundancy and complementarity. Based on these findings, we propose High-Rank Corrected Multi-Head Self-Attention (HR-MSA) to capture precise dependency information by high-rank attention maps without introducing additional computational burden. Our HR-MSA first utilizes the complete information of each pixel to compute an unabridged high-rank dependency. Then it independently applies linear corrections to different heads and achieves a high-rank weighted pixel information. Building around HR-MSA, we design a new architecture called High-Rank Attention for Super Resolution (HiRA-SR). Specifically, We develop Focusing Block (FB) to divert local pixel information from the HR-MSA module and introduce Residual Multi-Head Contextual Block (RMCB) to integrate global information through non-local attention. Experiments demonstrate that our HR-MSA can replace MSA and achieve efficient and effective improvements across various state-of-the-art SR methods. With parameters and FLOPs similar to SwinIR-light, our HiRA-SR sets a new state-of-the-art for lightweight image super-resolution. Our code will be available at: https://github.com/yyexplorerNB/HiRA-SR.