Visual Garment Collar Style Identification using Deep Neural Networks

Modeling fashion style is key for fashion industry. While traditional manually analysis is slow and inefficient, we aim to automate it using deep neural networks. Particularly, in this paper, we propose a VGG16-based collar recognition network that is more suitable for commercial use to solve the problem of difficult collar recognition in the apparel field. Furthermore, to enable our research and further research using deep learning, we create a dataset called COLLAR 2000. To achieve the commercial network for recognition accuracy, network stability, network complexity and other requirements, we explored traditional machine learning methods as well as mature neural networks, respectively, by adjusting the parameters of PCA, SVM, LeNet, AlexNet, VGG16 and combining migration learning methods to make them more suitable for garment collar recognition. The experimental results show that the VGG16-based collar recognition network is the best for collar recognition (81.7% recognition accuracy, 83.0%precision, 81.7% recall, 0.82 F1-Score), the network stability and complexity are moderate, which is more suitable for commercial use.