The subject of the article's research is the improvement of structural methods of image classification in computer vision systems. The goal is to reduce computational costs for classification by implementing a device for decomposing image description components using a system of orthogonal functions and implementing feature space compression models. Applied methods: ORB key point detector, set theory apparatus and vector spaces, metric models for determining relevance to sets of multidimensional vectors, theory of orthogonal decomposition of vectors, elements of probability theory, software modeling. Obtained results: modifications of the image classification method based on the introduction of orthogonal data decomposition in vector space were developed, models were proposed for data compression in the transformed feature space, Tanimoto metric was introduced for image comparison, a threshold selection method was established for determining equivalent description components. The effectiveness of the developed modifications of the classifier depends on the selection of a subset of functions for decomposition, the metric for comparing descriptions, and the method of determining the equivalence threshold. The implementation of the apparatus of orthogonal functions not only reduced computational costs tenfold, but also ensured sufficiently high indicators of classification performance and interference resistance. The practical significance of the work is the construction of new models of the image classifier in the transformed space of features, confirmation of the functionality, speed and immunity of the proposed modifications on examples of images, the creation of a software application for the implementation of the developed classification methods in computer vision systems.
Read full abstract