Blind Super-Resolution With Enhanced Features Integrating State Space and Self-Attention

CNN-based and Swin Transformer-based approaches have shown promising results in blind super-resolution. However, these methods often struggle to restore fine details accurately and simultaneously fail to strike an effective balance between performance and efficiency. To address these challenges, this work proposes the Degradation Awareness with the State Space Model and Transformer (DST) model with enhanced features, which balances global information extraction, computational efficiency, and detail restoration. The proposed model introduces a novel Residual Vision Mamba and Transformer Block (RVMTB), which synergistically combines the Cross Scan mechanism of the Vision State Space Model (VSSM) with the Shifted Window mechanism of Multi-Head Self-Attention (MHSA). This design establishes a global receptive field while sequentially reducing parameters and computational overhead, employing the VSSM and MHSA modules. Furthermore, a Multi-Branch Feature Enhancement (MBFE) module effectively integrates local and global features, while a contrastive learning-based degradation estimation module enables robust feature extraction under unknown degradations. Experimental results validate the superiority of the proposed method across benchmarks and scaling factors. Specifically, our DST model achieves PSNR gains of 0.33–1.01 dB and 0.08–0.21 dB over DASR and DSAT on Urban100 at ×4 scale, while maintaining higher efficiency with only 10.17 M parameters and 111.02 ms inference time compared to DSAT's 15.64 M and 161.07 ms, thereby establishing state-of-the-art performance.