On Neural Network Methods of Image Reconstruction and Super-Resolution

The methods for solving the problems of image inpainting and image super-resolution by means of image generation with neural networks are considered. Generative and adversarial neural networks are created and trained to solve them. Different variants of the structure of residual network blocks to increase the resolution are tested. The dependences of the recovery error on the fraction of damaged pixels, the quality of resolution increase on the number of residual blocks, and the choice of the training error function were obtained. It is shown that, in a wide range, the recovery quality almost does not depend on the fraction of damaged pixels, that adding residual blocks does not lead to its improvement, and that the generative adversarial network for resolution increase gives better results than the bicubic interpolation.