Image Classification in Fashion Domain

Abstract

Thanks to the wide spread usage of mobile devices with imaging capabilities, the number of digital images has been increasing exponentially. Every year, billions of images are taken worldwide. However, as the number increases, managing and arranging the pictures manually also becomes infeasible. Image classification is an elegant solution for this problem where one can define categories according to his/her needs thus, the images can be automatically separated and put into the categories. The flexible nature of image classification can make it possible to use it in different domains. In this thesis, it is shown that image classification can be also used in a specific field of fashion domain where the categories become more abstract. Three different approaches for solving the four categories image classification problem are proposed where the categories are “Informal Woman, Formal Woman, Informal Man and Formal Man”. Formality of a cloth is defined as the Western type of evening dresses where the dark colors with homogenous color distributions dominate the formal cloths and lighter colors with heterogeneous distributions. In order to bridge the gap between the pixels and the meaningful concepts such as color, shape and texture, four different low-level feature descriptors are used. For the shape, Histogram of Oriented Gradients is used whereas Local Binary Patterns is adopted for the texture description. The color distributions are described by using Color Structure Descriptor and Color Layout Descriptor. As the building block for the proposed classification topology, Support Vector Machines are used. The proposed approaches differ from each other in the way the features are extracted. Features are extracted globally in the Global Approach. In the second and third approaches, the images are divided into nine non-overlapping grids and the features are extracted separately from each of the individual grids. After the extraction process, they are concatenated into a single vector to represent the whole image. In the Seven Grids Feature Extraction Approach, the grids next to the head are excluded in order to further reduce the noise present. Once the features are extracted and concatenated, they are fed into a two-level classification scheme in which the male-female separation is followed by the formality check for the cloth. The experimental results show that Seven Grids Feature Extraction Approach outperforms the other two approaches yet the Global Approach shows close results.