A Study of Impression Term Extraction Method for Clothing Using Convolutional Neural Networks - For the purpose of supporting the visually impaired people-

Abstract

The visually impaired have difficulty in recognizing clothes that are very similar in shape. Since features and impressions of clothes are important in selecting clothes, this study examined a method for extracting impression words of clothes using deep learning. Ito et al.[1] used VGG16 to create a multi-class classification model that extracts two impression words from 16 impression words and a two-class classification model focusing on the impression word "casual,"In this study, we created a 2-class classification model for words other than "casual. As a result, a large difference was observed in the fit rate and the reproducibility of each impression word. The possible reasons for this are that the current impression words are selected based on color impressions and are not suitable as impression words for clothes, and that VGG16 is not suitable as a feature extractor for clothes. Therefore, we recreated the dataset using personal designs of fashion images. We also created a model in which VGG16 is replaced by EfficientNet. As a result, the evaluation value of the entire model was improved, and it became possible to perform one impression word with high identification accuracy. However, the 2-class classification can only focus on a single impression word. However, since the accuracy varies depending on the impression word, it is necessary to improve the accuracy of the 2-class classification by scaling up the model and changing the classifier.