Pixel-level Motion Deblurring Based on Multi-Scale Recurrent Networks
Zixu Tao, Di Lin, Jikang Mo, Yu Tang · 2024
Motion deblurring remains a significant challenge in the field of computer vision, with existing methods often struggling to fully restore the sharpness of edge features, resulting in noticeable visual artifacts. In this paper, we propose a novel pixel-level image deblurring method that leverages recurrent recursive networks (RRNs) to address these shortcomings. Our approach is built upon the Self-Residual Network (SRN) architecture, which has demonstrated strong performance in image restoration tasks. We enhance the SRN architecture by introducing a multi-scale fusion mechanism that effectively captures and integrates features at different resolutions, thereby improving the recovery of fine details, particularly in edge regions. Additionally, we streamline the network structure and reduce the number of parameters, making the network more computationally efficient and easier to train. This optimization also enhances the network's capacity to model and fit complex image features, leading to improved deblurring performance. Extensive experiments demonstrate that our method outperforms existing approaches in both quantitative metrics and visual quality, offering a promising solution to the persistent problem of motion blur in images.