The Evaluation of Text String Matching Algorithms as an Aid to Image Search

Abstract

The main goal of this paper is to analyse intelligent text string matching methods (like fuzzy sets and relations) and evaluate their usefulness for image search. The present study examines the ability of different algorithms to handle multi-word and multi-sentence queries. Eight different similarity measures (N-gram, Levenshtein distance, Jaro coefficient, Dice coefficient, Overlap coeffiient, Euclidean distance, Cosine similarity and Jaccard similarity) are employed to analyse the algorithms in terms of time complexity and accuracy of results. The outcomes are used to develop a hierarchy of methods, illustrating their usefulness to image search. The search response time increases significantly in the case of data sets containing several thousand images. The findings indicate that the analysed algorithms do not fulfil the response-time requirements of professional applications. Due to its limitations, the proposed system should be considered only as an illustration of a novel solution with further development perspectives. The use of Polish as the language of experiments affects the accuracy of measures. This limitation seems to be easy to overcome in the case of languages with simpler grammar rules (e.g. English).