Second International Workshop on Symbolic-Neural Learning (SNL-2018)

July 5-6, 2018
Nagoya Congress Center (Nagoya, Japan)

Deep Back-Projection Networks for Super-Resolution

Muhammad Haris (TTI), Greg Shakhnarovich (TTIC), and Norimichi Ukita (TTI)

Abstract:

The trends of deep super-resolution networks focus to propagate the low-resolution features in feed-forward architectures. The feed-forward network learns the input representations then plots it non-linearly to the high-resolution spaces. However, this approach does not explicitly show dependencies of low- and high-resolution images. Instead of predicting super-resolution image in such a feed-forward manner, we propose Deep Back-Projection Networks that exploit iterative up- and down-sampling layers with error feedback from projection error in each depth. We construct mutually-connected up- and down-sampling layers each of which represents different types of image degradation and high-resolution components. This deep concatenation of up- and down-sampling layers allows us to reconstruct a huge variety of super-resolution features and enable large scaling factors such as 8x enlargement. Experimental results show the effectiveness of our proposed networks compares to the state-of-the-art methods across multiple data sets.