Accuracy Improvement of Fashion Style Estimation with Attention Control of a Classifier

Computer vision has begun to be applied to fashion applications owing to the remarkable development of image recognition technology and deep learning. Fashion style estimation, the focus of this paper, is a hot topic as it provides useful information and services to consumers. This paper aims to enhance the accuracy of fashion style estimation by improving the classifier. To control the attention of a classifier, two different approaches are followed: human detection after segmentation with the aim of improving on previous schemes and enhancing the visual attention model (VAM). Experiments were conducted using the Hipster Wars dataset to train a final SVM-based classifier and the WEARStyle dataset to construct a feature extractor. The results using SSD-based detection after PSPNet-based segmentation, VAM with ICNet, and VAM with PSPNet showed classification accuracies of 83.8%, 84.7%, and 85.1%, respectively. All these outperformed VAM with full convolutional network (FCN), which resulted in a classification accuracy of 83.0% under the same experimental conditions. The results indicate that classification accuracy can be improved upon when the attention of the classifier is controlled.