The clothing image classification algorithm based on the improved Xception model

Abstract

In recent years, the clothing industry develops rapidly under the great influence of internet. A large number of clothing images have been produced and how to accurately make a classification of such a wide range of clothing has become a research focus. In this paper, we propose a clothing image classification algorithm based on the improved Xception model. Firstly, the last fully connected layer of the original network is replaced with another fully connected layer to recognise eight classes instead of 1,000 classes. Secondly, the activation function we employ in our network adopts both exponential linear unit (ELU) and rectified linear unit (ReLU), which can improve the nonlinear and learning characteristics of the networks. Thirdly, in order to enhance the anti-disturbance capability of the network we employ a L2 regularisation method. Fourthly, we perform data augmentation on the training images to reduce over-fitting. Finally, the learning rate is set to zero in the layers of the first two modules of our network and we fine-tune the network. The experimental results show that the top-1 accuracy by the algorithm proposed in this paper is 92.19%, which is better than the state-of-the-art models of Inception-v3, Inception-ResNet-v2 and Xception.