Realimentação de relevância em buscas de imagem usando programação Genética

Abstract

Fashion products are difficult items to be annotated and described by text, making it necessary to use images to perform searches on web sites of e-commerce. Such products hold great visual appeal, in other words the presentation of images relating to them are factors that directly influence the buying decision of a customer. These facts justify the study of the use of CBIR (Content Based Image Retrieval) in this context, an area already well studied in the scientific community, but that still has several shortcomings, the main one being the problem of Semantic Gap . The use of features extracted from the image by an algorithm is still not effective enough in associate it with its meaning, which is reflected in the results of a search, affecting the customer satisfaction with the store. This study seeks to address the problem of Semantic Gap through Genetic Programming and Relevance Feedback, motivated by the good results reported in the literature concerning the use of such techniques. Experiments were performed with an image base extracted from web sites e-commerce, and we used two subsets of images as queries, where one has images with a uniform background (as do the images of the data set), and the other has images with noisy backgrouns (photography in general). We compared the use of Relevance Feedback for both subsets, and for each subset we compared the use of ranking functions learned with and without using feedback. As the result, the best cenery for both subsets is to use the ranking function learned without usinf RF. Using RF on the learning process of GP makes the individuals dependent of the feedback, worsening the answers of searches before the first interaction with the user, and making the learned function unable to capture the semantic of the original query.