Інформаційна технологія пошуку заданих фрагментів в архіві аудіозаписів із застосуванням kd-дерев

Abstract

The Dissertation is devoted to improving recall and precision of the search results and increasing the speed on the basis of information technology for search of the specified audio fragments in the archive of audio recordings using kd-tree. With the development of the Internet, digital media, multimedia technology and access to the large multimedia archives became necessary to create new ways of navigation and interaction with the audio archives. Despite the great progress achieved in this area in the recent years, modern information retrieval systems have some shortcomings, in particular: - the insufficient performance with regard to the large archives; - the impossibility of the direct application to solve such tasks as broadcast monitoring on the radio (television) to determine the repetition number of some audio recording (e.g. an advertising product, a musical piece, etc.) during translation. The offered model of the audio body, which contains an array of audiofiles, a database of parameters and metadata, and dynamic kd-tree, allowed to reduce the audio fragment duration, selected for the search, and implement information technology for search of audio recordings with the different duration. It is shown that the lowest duration of the audio fragment that ensures the recall of the search results at 99.4%, should be 1s. The method of k-means clustering got the further development, differs from the existing ones by the improved procedure of selecting the vector as a new centroid by successive run of k-means that provides a solution, close to the global minimum of the clustering error. The search method based on kd-tree is improved, which unlike the existing, uses the improved estimation of the measure proximity, enhancing the rallback and precision of search results. The method of the combined search of audio recording for a specified audio fragment is proposed, based on the implementation of the fast approximate kd-tree search of the several nearest recordings in the parameter database of the reduced dimension on the first stage of the search, among them the relevant audio recording is defined on the second stage, which resulted in reducing the time search compared to an accurate kd-tree search.The proposed methods, as well as algorithms and software, developed on their basis, allowed to search the specified audio fragment in real time. The research results implemented at the Institute of Electronics and Communications of the Ukrainian Academy of Sciences and in the educational process of VNTU at the course "Digital Signal Processing".