Research on multi-label clothing image recognition based on deep learning

Purpose Nowadays, the labeling of clothing images on e-commerce platforms mainly relies on manual labor, which not only wastes manpower and time, but also causes problems such as inconsistent labeling standards among merchants and cognitive differences among different annotators. Besides, it is particularly difficult to make image annotation for dresses with rich style variations and massive image data. To solve this problem, in this study, an automatic recognition method of multi-label images based on deep learning is proposed. Design/methodology/approach In this paper, first dress images were selected whose attributes were annotated and a dataset for the dressing images was established according to the style of silhouette, collar, and sleeve. Then based on the baseline networks, a multi-label dress image recognition method, the VE-DE network, was proposed, which combines VGG-19, DenseNet-201 and the channel attention network. Findings The results show that the average precision mean (mAP), precision, and recall of the dress images in the VE-DE recognition test set were 93.9%, 91.6%, and 90.3%, respectively, all higher than the three baseline networks. Among them, mAP was 4.3%, 11.1%, and 3.7% higher than VGG-19, ResNet-101, and DenseNet-201, respectively. VE-DE network can improve the vulnerability of the baseline network to the influence of external factors when identifying the profile and has a good recognition effect on the samples with easily confused profiles. Visual analysis of the class activation heatmap and intermediate layer activation of the VE-DE model shows that it can focus on such parts of the dress as the collar, sleeves, and waist, which contribute significantly to the recognition of the images, with the waist being the most highly focused area. Originality/value The automatic recognition method of multi-label images can be used in an e-commerce platform. With the help of this technology, e-commerce platforms can automatically and quickly label clothing images, reduce labor costs, improve product labeling speed and quality.