Multi-Scale Image Deblurring Using Wavelet Transform and Attention-Based Feature Fusion

Image deblurring is a classical problem in low-level computer vision, aiming to reconstruct high-quality sharp frames from blurred images. Recently, deep learning-based image deblurring methods have significantly surpassed traditional algorithms in performance. In this paper, we propose an image deblurring method using wavelet transform and attention mechanisms. Our approach builds upon the existing Multi-Input Multi-Output U-Net (MIMO-UNet) architecture, which demonstrates excellent performance in deblur-ring tasks. By introducing wavelet transform into deep residual networks, we inject frequency information into the network, enabling it to use high-frequency components for more precise capture of edge and texture details. Furthermore, we use attention mechanisms in the feature fusion module to achieve channel-wise adaptive fusion of multi-scale information, while introducing spatial attention to refine the fused feature maps, further enhancing image restoration performance. Extensive quantitative experiments and qualitative evaluations demonstrate that our proposed method surpasses existing approaches in terms of PSNR and SSIM metrics.