Kansei Retrieval of Clothing using Features Extracted by Deep Neural Network

Abstract

In this paper, we construct Kansei retrieval of clothing. This system is designed to search for user's clothing preference by considering their Kansei. Then, this system uses features of clothing images. In previous study, the features were 20 bits and extracted by image processing. However, these features aren't enough for Kansei retrieval. So, our proposal is Kansei retrieval of clothing on the basis of the features extracted automatically. This system is constructed with two neural networks. The first neural network is deep neural network that extracts the features. The second neural network identifies user's evaluation of clothing using the features. In this study, we verified the effectiveness of this system by a simulation. Result of the simulation showed this system was superior to the previous study in 46.65% increasing accuracy of imitating user's Kansei and 30.43% decreasing error of Kansei retrieval. Thus, the effectiveness of the proposed system was confirmed.