Clothing fashion style recognition with design issue graph
- PDF / 2,404,420 Bytes
- 13 Pages / 595.224 x 790.955 pts Page_size
- 17 Downloads / 194 Views
Clothing fashion style recognition with design issue graph Xiaodong Yue1
· Cheng Zhang1 · Hamido Fujita2 · Ying Lv1
Accepted: 14 September 2020 © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Fashion style recognition of clothing images facilitates the clothing retrieval and recommendation in E-commerce. It is still a challenging task because the clothing images of same style may have diverse visual appearances. Existing fashion style recognition methods utilize deep neural networks to classify clothing images based on pixel-level or region-level features. However, these features of local regions lack the semantics of fashion issues and make the style recognition sensitive to clothing appearance changing. To tackle this problem, we construct Design Issue Graphs (DIGs) with clothing attributes to form global and semantic representations of fashion styles, and propose a joint fashion style recognition model which consists of two convolutional neural networks based on clothing images and DIGs. The experiments on DeepFashion data sets validate that the proposed model is effective to recognize the clothing fashion styles of diverse appearances. The integration of DIGs into Deep Convolutional Neural Networks (DCNNs) achieves 1.75%, 0.99%, 1.03%, 1.53% improvements for multi-style recognition and 1.22%, 2.06%, 1.58%, 2.20% improvements for certain style recognition in the evaluations of accuracy, precision, recall and F1-score on average respectively. Keywords Fashion style recognition · Design issue graphs · Convolutional neural networks
1 Introduction With the boom of E-commerce, fashion style recognition of clothing images has attracted lots of research interests because of its potential values to clothing retrieval [3, 10, 17, 20, 21] and recommendation [12, 18, 29, 41]. However, it is still a challenging task because the clothing images of same style may have diverse visual appearances. For example, the clothing images which have diverse textures, colors and fabric features may belong to the same fashion style. Moreover, the clothing items in the images frequently subject to deformations and occlusions, which also makes the recognition of fashion styles difficult. Xiaodong Yue
[email protected] Hamido Fujita [email protected] 1
School of Computer Engineering and Science, Shanghai Institute for Advanced Communication and Data Science, Shanghai University, Shanghai, China
2
Faculty of Software and Information Science, Iwate Prefectural University, Iwate, Japan
Most of the existing fashion style recognition methods classify clothing images based on the pixel-level or regionlevel features [1, 48, 52]. However, fashion styles are actually semantic descriptions of clothes so that highlevel semantic features are essential for accurate and robust recognition of fashion styles. The features of local regions lack the semantics of fashion issues and make the style recognition methods sensitive to clothing appearance changing [16, 26, 28]. Therefore, to achieve the accurate and robus
Data Loading...
