Fashion Style Recognition with Graph-Based Deep Convolutional Neural Networks

Abstract

Recognizing fashion styles of clothing from images plays an important role in the application scenarios of clothing retrieval and recommendation in E-commerce. Most existing works directly utilize the machine learning methods such as Deep Convolutional Neural Network (DCNN) to classify clothing images into different styles. However, these image classification methods are totally data-driven and neglect the domain issues of apparel fashion design. To tackle this problem, we propose a domain-driven clothing style recognition method in this paper, which involves both image classification and domain knowledge of fashion design. Specifically, we formulate the domain knowledge of design elements with the undirected graphs of clothing attributes and thereby build up a domain-driven fashion style classifier with Graph-Based DCNN. Synthesizing the classifications based on both clothing images and the graphs of design elements, we produce the final clothing style recognition results. The experiments on Deep Fashion database validate that the proposed clothing style recognition method can achieve more precise results than the traditional data-driven image classification methods.