Image Super-Resolution with Deep Dictionary

Since the first success of Dong et al., deep-learning-based approach has become dominant in field single-image super-resolution. This replaces all handcrafted image processing steps traditional sparse-coding-based methods with a deep neural network. In contrast to methods, which explicitly create high/low-resolution dictionaries, dictionaries are implicitly acquired as nonlinear combination multiple convolutions. One disadvantage is that their performance degraded for images created differently from training dataset (out-of-domain images). We propose an end-to-end super-resolution network dictionary (SRDD), where high-resolution learned without sacrificing advantages learning. Extensive experiments show explicit learning makes more robust out-of-domain test while maintaining in-domain images. Code available at https://github.com/shuntama/srdd .