Cross-Stage Patch Interaction for Coarse-to-Fine Image Deblurring

Source abstract

Real-world motion blur exhibits complex, spatially varying characteristics that challenge existing deblurring methods. While multi-stage architectures offer progressive refinement, simple inter-stage interactions often limit effective information transfer. To address this, we propose CPINet, a Cross-stage Patch Interaction Network for coarse-guided image deblurring. CPINet employs a two-stage coarse-to-fine framework: the first stage uses an Uncertain Flow-based Prior Network (UFPNet) for coarse restoration, followed by a second stage for fine-grained patch-level refinement. The second stage utilizes a Coarse-guided Patch Encoding (CPE) block with a Patch-wise Cross-Channel Attention Gating (PCAG) mechanism to integrate first-stage features. Furthermore, a Patch-based Multi-Kernel Dual-Branch (P-MKDB) block is incorporated into the decoder to effectively handle diverse local blur patterns. Experiments on GoPro, HIDE, and RealBlur show competitive deblurring performance, while ablation studies support the contribution of the proposed cross-stage patch interaction and patch-wise refinement. Notably, CPINet improves the mean PSNR from 36.98 dB to 37.61 dB and the mean SSIM from 0.954 to 0.961 relative to UFPNet, with both metrics averaged across RealBlur-J and RealBlur-R. These improvements demonstrate its effectiveness in challenging real-world scenarios.