Channel Graph Convolutional Networks for Animation Image Super-Resolution

Animation images exhibit rich edge characteristics, distinct from those found in naturalistic images. Since general image super-resolution reconstruction techniques are not explicitly designed to handle animation image features, especially edges, these methods may not be suitable for animation image reconstruction. To address this limitation, we have curated a large-scale, high-quality color animation dataset encompassing various contents and styles to investigate this phenomenon. In response to the unique challenges posed by animation images, we propose Channel Graph Convolutional Networks (CGC-Nets), which are fast, accurate, and lightweight for animated image super-resolution. These networks demonstrate efficient memory usage and runtime performance across different scaling factors. Specifically, CGC-Nets consist of two dedicated sections. The Back-Projection block is used in the first section to maintain and modify the animation picture content (i.e., enhance the animation image’s low-frequency information). The second step focuses on improving the edge information of animated images by transforming feature maps obtained from gradient maps into graph features. Graph features encapsulate the topological relationships between each channel in the feature generated by convolution. This structuring of details and edge information (i.e., high-frequency information) in animation images makes it easier to capture and reconstruct such information effectively. Our proposed CGC-Nets method significantly enhances the performance of anime image super-resolution tasks. Extensive experiments validate the superiority of the CGC-Nets network over state-of-the-art methods in super-resolution reconstruction.