Recognition of low-resolution building images using deep learning and softmax

Abstract

Image recognition of buildings has been gaining attention in recent years due to the growing popularity of georeferencing tools and their use in the areas of tourism and navigation. The use of mobile devices is critical in these applications and, therefore, processing should be fast and allow the use of low-resolution images. However, many options to perform this task use highly complex recognition techniques. Thus, the great difficulty is that these images are obtained from different angles, under different lighting conditions, in addition to partial obstructions from trees, moving vehicles, or other buildings that hinder recognition. To bypass these problems, this work aims to use the Deep Stacked Autoencoder in the extraction and reduction of features, the Softmax classifier in the recognition and the same deep learning technique applied to the classifier that uses the multilayer perceptron network for comparison purposes with Softmax software. To verify the efficiency of the method, a database of images of buildings from the University of Shefield was used. The algorithm proved to be efficient in the recognition task, reaching 94.9% accuracy.