Automated Object Recognition System based on Convolutional Autoencoder

This paper describes the model, implementation and the experimental verification of an aerial image processing and recognition technology based on artificial neural networks, specifically, convolutional autoencoders and classifying perceptrons. An originally developed model that includes autoencoder preprocessing for compression and extraction of informative features was applied to the task of pattern recognition, namely, locating and identifying the objects of certain higher-level classes of interest in the images produced by aerial photography. Classification efficiency of the method was measured and compared with other common methods of classification, the advantages and shortcomings of the proposed approach analyzed and potential applications in real-time remote object recognition systems, as well as in automating the generation of training data for image recognition systems discussed.