Shirt Semantic Classification Based on Convolutional Neural Network

Abstract

With the rapid development of e-commerce, a large number of clothing images have flooded into the Internet replacing the keyword search with a graph search has become a new trend. However, the inaccurate semantic classification of clothing leads to the low accuracy of graph searching and the poor retrieval effect. In order to solve this problem, this paper proposes a parallel shirt-style classification algorithm which greatly improves the accuracy. This paper mainly completed the following work: Firstly, the shirt sample library was established by means of questionnaire survey and expert voting. Then, the mapping mechanism was used to extract the ROI area and construct the shirt-Attention branch network. Last, the backbone is combined with the branch to make up the ShirtNet. The experimental results show that the introduction of parallel network makes the classification accuracy rate up to 91% which is much improved compared with traditional machine learning and convolutional neural networks.