MLRN: A multi-view local reconstruction network for single image restoration

Abstract

Limited by storage conditions, the degradation of old photos exhibits complex and diverse features. Existing image restoration methods heavily rely on features extracted from a single view, allowing them to effectively handle a certain type of noise; however, they lack the versatility to address multiple types of noise. This poses great challenges in accurately detecting various noises and integrating both global and local characteristics. To solve this problem, we propose a multi-view local reconstruction network (MLRN) for repairing old photos. Specifically, we first conduct multi-granular semantic-based segmentation on the defaced images. Then, we extract and fuse features from different views (granularities) to accurately identify multi-type noises. Afterwards, based on these different views, the image blocks are reconstructed multiple times through fast Fourier convolution. Finally, an attention mechanism is incorporated to correct and fuse the image blocks, achieving optimal performance from both global and local perspectives. Experiments conducted on two image datasets show that the proposed model has better performance than the baseline methods in qualitative and quantitative comparisons. It outperforms the state-of-the-art approaches on average by 16%, 15%, 18% and 10% in terms of PSNR, SSIM, FID and LPIPS, respectively, highlighting the significance of multi-view features in old photo restoration tasks.