SSHRF-GAN: Spatial-Spectral Joint High Receptive Field GAN for Old Photo Restoration

Old photo restoration is a challenging problem since it suffers from mixed degradation like noise, blurriness, scratch, dots, or color fading during its preserving or digitalizing process. The common practice of old photo restoration is to apply single degradation restoring methods sequentially, but it cannot restore the old images at one time. The existing mixed degradation restoring methods are designed in a complex architecture or training paradigm. To tackle this problem in a simple way, we propose SSHRF-GAN: Spatial-Spectral joint High Receptive Field GAN. It contains three independent modules for the mixed degradation that appeared in old photos: (1) High Receptive Field Fast Convolution block for noise and blurriness restoration; (2) Partial convolution block for scratch, dots, and cracks inpainting and (3) Image-wise histogram attention for color tone mapping. We evaluate our methods on synthesized datasets and gather real old data. The result shows that our simple network has a comparative performance with those sequential applied single degradation restoration models and previously proposed complex models.