Processing of Incomplete Images by (graph) Convolutional Neural Networks

Abstract

We investigate the problem of processing incomplete images by neural networks without replacing missing values. To deal with this problem, we first represent an image as a graph, in which missing pixels are entirely ignored. The graph image representation is processed using Geo-GCN -- a type of graph convolutional neural networks, which is a proper generalization of classical CNNs operating on images. On one hand, our approach avoids the problem of missing data imputation while, on the other hand, there is a natural correspondence between CNNs and Geo-GCN. Experiments confirm that our approach performs better than analogical CNNs with the imputation of missing values on typical classification and reconstruction tasks.