EAAI Journal 2025 Journal Article
Lightweight multi-level feature integration transformer for image super-resolution
- Shuheng Wang
- Ziao Gong
- Mengda Li
- Shen Wu
- Yilin He
Transformer-based methods have attracted significant attention in the field of image super-resolution. However, existing approaches typically concentrate on a single level of information for image reconstruction, and neglect the critical role of multi-level information in feature representation, resulting in the low utilization of the potential capabilities of Transformer. To address this limitation, we propose a novel Transformer model, named as Multi-Level Differential Window Attention Transformer (MLDAT), designed for multi-level information fusion. Specifically, our approach introduces a self-attention module based on differential windows to comprehensively extract and integrate feature information across varying window sizes. Additionally, we introduce a High-order Global Attention Module (HGAB) to combine the second-order attention with global self-attention, which facilitates the establishment of relationships between local features and the overall global feature context within an image while complementing local window information. Extensive experimental results demonstrate that our model can significantly improve the performance of image super-resolution, and achieves the better results compared with existing methods.