Development of an object-oriented image comparison algorithm for efficient search

• Published in: Sučasnij stan naukovih doslìdženʹ ta tehnologìj v promislovostì (Online) no. 2(32); pp. 79 - 101
• Main Authors: Prokopenko, Oleksandr, Smelyakov, Serhii
• Format: Journal Article
• Language: English
• Published: 30.06.2025
• ISSN: 2522-9818; 2524-2296
• DOI: 10.30837/2522-9818.2025.2.079

The object of research is content-based image retrieval (CBIR). The subject of this study is models and methods for content-based image retrieval (CBIR) and managing large volumes of media content in extensive image storage systems. The goal of the research is to develop an algorithm for comparing object-oriented image descriptors, which involves using advanced computer vision models for object detection and constructing efficient methods for comparing and searching these descriptors. The proposed descriptor and comparison algorithm aim to enhance the efficiency and accuracy of image search and management processes. The tasks include: analyzing modern approaches and solutions for creating and comparing image descriptors and their use in CBIR; developing metrics and algorithms for comparing image descriptors that effectively utilize information about detected objects – such as their types, sizes, and locations – for image search in large data repositories; conducting experiments to evaluate the proposed image search algorithm and comparing its efficiency with existing solutions. The methodology includes: conducting a comprehensive review of advanced image descriptor generation methods, including hash-based descriptors, handcrafted descriptors, and deep learning-based descriptors; analyzing the use of existing descriptors in CBIR systems, focusing on their advantages and limitations; evaluating the best image search algorithms, including deep learning-based approaches; developing an object descriptor comparison algorithm for tag-based search, image-based search, and other tasks. The results obtained are as follows: an object-based image descriptor was developed using state-of-the-art machine learning models for object detection; metrics and comparison algorithms for the proposed descriptors were developed, enabling their use for CBIR in large data repositories; a series of experiments were conducted to assess the efficiency and search quality of the proposed descriptor and algorithms in large-scale image storage systems. These experiments compared their performance with existing methods, revealing their advantages and limitations, namely: faster descriptor generation; faster descriptor comparison than hashed, handcrafted, and deep learning-based descriptors; efficient image filtering in storage; higher search quality and speed for image-based queries. However, the descriptor’s effectiveness depends on the quality of the model and data used for object detection, as images without detected objects do not appear in search results, which may limit search completeness. Conclusions: The developed algorithm for comparing object-oriented image descriptors is an effective tool for solving various CBIR tasks. The obtained results are satisfactory, as the proposed image search algorithm outperforms most alternatives in terms of speed and search quality. A promising direction for future research is the development of a CBIR system using the proposed descriptor and algorithms, enhanced by parallel and distributed computing, and further refinement for specific applications. This would allow its use not only for general-purpose images but also for more precise scientific domains.