The Study of Fashion Style Classification: Harajuku-type Kawaii and Street Fashion

Abstract

This work provides a comparison among four fashion image classification models, which are a random forest classification method, a convolutional neural network (CNN), a CNN model trained with preprocessed images that a human head and face are removed, and a CNN model trained with images that a head, face, and background are removed. We are interested in three fashion categories, including Harajuku kawaii style, Thai street style, and European street fashion styles. We retrieved images from Google Images, and adopted various preprocessing techniques, which were human detection, clothing segmentation, landmark detection, and colorfulness matrix to locate humans and clothes, and extract fashion features. We grouped the images into 2 classes, which were Harajuku kawaii fashion and other street fashion styles. Results showed that the accuracy score from the CNN model without cropping and background removing process was greater than the others. However, when we validated the models using the fashion dataset that was evaluated by fashion experts, the result indicated that the performance of the CNN model trained with the head-cropped image dataset was greater than the others. We concluded that the deep learning model trained with the head-cropped dataset achieved better results because it learned only clothing features, which are the most significant information in the fashion domain.