ANALYSING AERIAL PHOTOGRAPHS USING A CONVOLUTIONAL AUTOENCODER NETWORK

Abstract

The purpose of this work was to create a functioning autonomous system capable of automatically detecting anomalies in aerial images obtained using unmanned aerial vehicles. Anomalous, in this context, are objects and segments of the image that do not fit with the general picture of the investigated area, such as man-made objects and vehicles in the wilderness, people in restricted areas, etc. The research considers the informative capabilities of convolutional neural networks (CNN) for solving the task of detecting abnormal objects in photographs. The structure of convolutional autoencoder with three encoding and three decoding convolutional neural layers is proposed, providing the capability of learning on unlabeled images and recognizing previously unknown types of anomalies. In addition, we study the impact of applying preprocessing algorithms on the speed and effectiveness of the system. Such algorithms include detection of color and brightness threshold, as well as finding the contours of objects that stand out in the image. We also propose a method of localizing potentially abnormal segments using the combination of these algorithms. We show that this approach allows for a great increase in performance at the cost of a very marginal increase in the number of anomalies missed.

The resulting system consists of a convolutional autoencoder model, two preprocessing algorithms based on detecting thresholds of color and brightness, an algorithm for contour detection and a final classifier, weighting the outputs. The developed NN model was trained on several data sets containing aerial photography images with no anomalies. The resulting system was tested out on real-world data, and the results of this testing are provided here.