Terrain Image Recognition with Unsupervised Generative Representations: the Effect of Anomalies

Abstract

Recognition of real-world images plays important role in a rapidly expanding range of applications. In this work we examined characteristics of distributions of classes of aerial images of terrain obtained in real time aerial photography from the perspective of detection and rectification of outliers in the generative latent distributions of image classes. Neural network models of generative self-learning with the architecture of convolutional autoencoder were used for compression of image data and extraction of informative features via selection of most informative principal components. Influence of outliers was examined with respect to distributions of image classes in the informative low-dimensional latent representations and classification performance of models trained in such representations. It was found that removal of outliers results in significantly more compact latent distributions of characteristic types of images, with a positive impact on classification performance. The results can be used in developing methods of learning based on unsupervised generative structure in informative representations in applications and problems with a deficit of training data.