Image deblurring processing model based on dual-pixel image imaging

Dual-pixel imaging introduces sub-pixel response differences and spatially non-uniform blur, making it difficult for conventional methods to preserve both local details and global structural consistency. This study proposes a dual-pixel image deblurring model integrating feature alignment and fusion, a multi-scale UNet, and Transformer-based global modelling. Shared-weight feature extraction and adaptive fusion are employed to exploit complementary information from left and right sub-pixel images. The multi-scale UNet reconstructs local details at different resolutions, while the Transformer embedded in the bottleneck layer captures long-range dependencies and enhances global structural consistency. Experiments on the GoPro and RealBlur datasets show that the proposed method achieves a peak signal-to-noise ratio of 33.9 dB, a structural similarity index of 0.967, and an inference speed of 45.2 FPS at 512 × 512 resolution. These results demonstrate that the proposed model provides an effective and efficient solution for high-quality restoration of complex non-uniformly blurred images.