Evaluation of Highly Compressed Semantic Features for Efficient Image Representation

Visual information means large amounts of data. Therefore, methods of representing images using sparse features are constantly being researched, which are effective in at least two aspects: they enable accurate classification and are lightweight. In the paper, we present a number of this type of sparse features, which arise from various combinations of features originating from the semantic layers of the ResNet-50 network and compressed with one of two dominant component analysis methods. The first is principal component analysis and the second is neighborhood component analysis. The obtained results confirm our initial assumptions that various combinations of features and component analysis enable both classification and content-based image retrieval at the level of the best performing methods.