Unsupervised Feature Extraction - A CNN-Based Approach

Working with large quantities of digital images can often lead to prohibitive computational challenges due to their massive number of pixels and high dimensionality. The extraction of compressed vectorial representations from images is therefore a task of vital importance in the field of computer vision. In this paper, we propose a new architecture for extracting such features from images in an unsupervised manner, which is based on convolutional neural networks. The model is referred to as the Unsupervised Convolutional Siamese Network (UCSN), and is trained to embed a set of images in a vector space, such that local distance structure in the space of images is approximately preserved. We compare the UCSN to several classical methods by using the extracted features as input to a classification system. Our results indicate that the UCSN produces vectorial representations that are suitable for classification purposes.