Uma abordagem estocástica para recuperação de imagens de roupa baseada em conceito utilizando o SVM e o feedback de relevância do usuário

Abstract

Image retrieval systems have become a common approach in online fashion applications, allowing consumers to search for clothing items. However, the task of exploring large data sets of images through naive retrieval methods is generally inefficient and tedious for users. When relevance feedback (RF) from the user is considered, machine learning methods may take advantage of user preferences to enhance the experience. Nevertheless, most RF approaches focus on retrieving only user relevant images, disregarding the learning curve of the machine and resulting in a lack of diversity. In particular, when a new user begins using the system, there is a lack of information about him, resulting in unsatisfactory recommendations (the cold start problem). We propose a machine learning approach based on RF to retrieve clothing images using a threefold strategy: retrieve user relevant images; retrieve images that do not comply with user learned preferences, to avoid convergence to local minimal; and retrieve images from uncertainty regions, to improve the learning curve, all in a stochastic manner. To mitigate the cold start problem, we present a novel sampling method to improve the diversity of retrieved images that employs a combination of a multidimensional projection method with an adaptive spatial data structure. Our approach is validated through quantitative and qualitative user experiments using an annotated clothing images data set to evaluate the effectiveness and efficiency of user-relevant image retrieval. Results revealed that our approach can rapidly improve the retrieval of appropriate images in a few user iterations while providing higher diversity than straightforward approaches.