Cross-Hierarchical Structure-Detail-Aware Transformer for Single Image Deblurring

Blurred images, resulting from a mix of camera shake and object motion, typically exhibit directional and uneven blurring, which diminishes overall visual quality. Despite many single-image deblurring methods proposed in recent years, their effectiveness remains limited, particularly in real-world scenarios involving different scales, degrees of depth, and background-object confusion. To address these issues, we propose a novel Cross-hierarchical Structure-Detail-aware transFormer (CSDFormer) for single image deblurring. Our model operates on multiple scales (multi-scale) and layers (multi-layer), focusing on both large-scale and local blur as well as varying blur depths. We introduce the Structure-aware Feature Extraction (SaFE) Module and Detail-aware Feature Extraction (DaFE) Module to extract significant features layer by layer across different scales (cross-layer). For effective feature exchange between different scales (cross-scale), we propose two cross-scale information exchange modules: the fine-to-coarse and coarse-to-fine Cross-scale LSTM Information Exchange (CsLSTM-IE) modules. These modules aim to restore structural and textural details by progressively learning features from fine to coarse scale and back. The experimental results demonstrate that the CSDFormer model outperforms state-of-the-art methods in both synthetic and real-world datasets, both quantitatively and qualitatively. It excels in removing blur across various scales and depths, restoring backgrounds, and preserving details.