Enhancing fashion style classification by capturing local details

The key issue in fashion style classification is analyzing local details, which are locally expressed characteristics of an outfit, in an image. In this paper, we propose Fashion-LD, a fashion style classification model with effective local detail capturing, which emphasizes the information on local details using an attention mechanism and a trainable global pooling method. An attention mechanism accentuates salient pixels in intermediate feature maps, and a trainable global pooling method emphasizes pixels containing important information when summarizing final feature maps into a feature vector. Specifically, global bivariate normal pooling method that exploits the spatial information of pixels is newly proposed and utilized in Fashion-LD. Results of the experiments conducted on K-Fashion dataset show that Fashion-LD outperformed other image classification models, and both the attention mechanism and the trainable global pooling method have shown to play a role in enhancing fashion style classification performances. Furthermore, we propose a visualization method for Fashion-LD to generate an intuitive visual explanation for a classification result of the model. Comparing visual explanations obtained by the proposed visualization method with those drawn by Grad-CAM suggests that the results of our visualization method had better interpretability than those of Grad-CAM.