Convolutional Neural Networks for Clothing Image Style Recognition

Abstract

Automatic recognition of the clothing image's style is important for quite a few applications, including apparel automatic labeling, recommendation for clothing, and clothing retrieval, etc. Convolutional neural networks cope with the image recognition well. However, the networks require a fixed-size input via cropping or scaling the image arbitrarily, which may reduce the recognition accuracy for the images. This paper equipped the fine-tuned VGG-Net with spatial pyramid pooling to eliminate the restriction of a fixed-size input image. The study showed that the combined network had a higher cross-validation accuracy of the style recognition in clothing images compared with the Google-Net and the fine-tuned VGG-Net. The network for the style recognition of clothing images flexibly addresses the issue of the dataset with different sizes and scales. This study also improves the accuracy of the style recognition in clothing images. Moreover, the network is beneficial to the classification or recognition of other datasets.