Export Reviews, Discussions, Author Feedback and Meta-Reviews

PAPER SUMMARY A recent work [4] proposed using a deep CNN for super-resolution. Although it generated impressive results, it could not outperform a state-of-the-art super-resolution method based on sparse coding [10]. The authors of this paper argue that this limitation is attributed to the inability of the CNN model in [4] to handle translation variant interpolation (TVI). To address this problem, they propose using an old interpolation method to develop a CNN variant called Shepard convolutional neural network (ShCNN). ShCNN is compared with other methods on inpainting and super-resolution tasks.