Image inpainting with gradient attention

We present a novel modification of context encoder loss function, which results in more accurate and plausible inpainting. For this purpose, we introduce gradient attention loss component of loss function, to suppress the common problem of inconsistency in shapes and edges between the inpainted region and its context. To this end, the mean absolute error is computed not only for the input and output images, but also for their derivatives. Therefore, model concentrates on areas with larger gradient, which are crucial for accurate reconstruction. The positive effects on inpainting results are observed both for fully-connected and fully-convolutional models tested on MNIST and CelebA datasets.