Image Recognition Using Data Distributions for Quantized Structural Description

The aim of this research is to improve the performance of structural image recognition methods based on descriptions in the form of sets of keypoint descriptors. The main focus is on the applying of quantization techniques and their characteristics to reduce computational costs in the task of image classification with a fixed set of reference images. As a result of feature space transformation, the reference image description base is compactly represented by a set of quantization centers, the use of which significantly accelerates classification. Each image in this approach can be represented as a vector of integers reflecting the number of elements in each quantum. The essence of the improvement is the use of a distribution matrix of reference data across the set of classes during classification. The application of distributions enhances alignment with reference information and ensures that classification accuracy remains at a high level while substantially speeding up analysis procedures. Several modifications of the classifier with both full and reduced representations of the distribution matrix, taking into account the significance and homogeneity of its components, were studied experimentally. The classification speed for the proposed improvement, compared to linear search, increases proportionally to the number of reference images and, for example, having 10 images, it achieves a speedup of more than 220 times. Experimental modeling results of the proposed methods are presented for images from two art painting datasets. The test set was formed as a collection of images from the reference image database and out-of-database images, with geometric transformations applied to them. The results show that the use of distributions ensures high classification accuracy with a significant gain in processing speed. The implementation of the considered methods is advisable in applied computer vision tasks with high requirements for the speed of visual object recognition.