基于服装风格的款式相似度算法

doi:10.13475/j.fzxb.20201207808

摘要/Abstract

摘要：

为准确计算服装图像之间的相似度,从而满足更多用户通过搜索服装搭配图像来购买相似服装的跨场景需求,研究了服装风格的影响因素,并以服装风格的量化标准为基础,构建了服装款式的风格特征模型,进一步分析了现有服装属性识别算法的不足,实现了基于深度学习的款式风格特征识别,通过构建融合迁移学习的残差网络模型,刻画出服装在款式上的风格特征。实验结果表明:模型在服装款式风格特征上的精确度接近90%,准确度达到了80%;相对于传统的图像相似度计算方法,基于服装款式风格的图像相似度计算,准确率和可解释性更高,也为服装个性化推荐提供了新的思路。

关键词: 服装风格, 图像相似度, 深度学习, 款式风格特征, 迁移学习, 残差神经网络

Abstract:

In order to calculate accurately the similarity between clothing images so as to meet the cross-scene needs of more users searching for clothing matching images for purchasing similar clothing, the influencing factors of clothing styles were investigated based on clothing styles constructed according to quantitative standards. The feature model of styles was established to further analyze the shortcomings of existing clothing attribute recognition algorithms, and was used to identify styles based on deep learning. The features of clothing styles were depicted through constructing a residual network model that integrates transfer learning. The experimental results show that the precision of the model on clothing style features is close to 90%, and the overall accuracy reaches 80%. Compared with the traditional image similarity methods, the accuracy of the image similarity calculation based on clothing styler is higher.This research also provides new ideas for personalized clothing recommendation.

Key words: clothing style, image similarity, deep learning, style characteristic, transfer learning, residual neural network

中图分类号:

TP391.41

江慧, 马彪. 基于服装风格的款式相似度算法[J]. 纺织学报, 2021, 42(11): 129-136.

JIANG Hui, MA Biao. Style similarity algorithm based on clothing style[J]. Journal of Textile Research, 2021, 42(11): 129-136.

图/表 10

图1

图2

图3

图4

图5

表1

ResNet50的网络结构"

网络层名称	输出大小/个	ResNet50网络结构
卷积层1	112×112	7×7, 64, 步长2
卷积层2	56×56	$1 × 1 64 3 × 3 64 1 × 1 256$ ×3
卷积层3	28×28	$1 × 1 128 3 × 3 128 1 × 1 512$ ×4
卷积层4	14×14	$1 × 1 256 3 × 3 256 1 × 1 1024$ ×6
卷积层5	7×7	$1 × 1 512 3 × 3 512 1 × 1 2048$ ×3
全连接层	1×1	平均池化,指数函数softmax

表1

图6

表2

图7

表3

参考文献 18

[1]	韩旭. 基于多特征融合的商标图像检索技术研究[D]. 郑州:郑州轻工业大学, 2020:42-43.
	HAN Xu. Research on trademark image retrieval technology based on multi-feature fusion[D]. Zhengzhou: Zhengzhou University of Light Industry, 2020:42-43.
[2]	王朝卿, 沈小林, 李磊. 图像相似度计算算法分析[J]. 现代电子技术, 2019, 42(9):31-34,38.
	WANG Chaoqing, SHEN Xiaolin, LI Lei. Analysis on image similarity calculation algorithm[J]. Modern Electronic Technology, 2019, 42(9):31-34,38.
[3]	丁维龙, 辛卫涛, 徐志福, 等. 基于图像特征的植物形态相似度算法[J]. 中国图象图形学报, 2019, 24(12):2255-2266.
	DING Weilong, XIN Weitao, XU Zhifu, et al. Plant morphology similarity algorithm based on image features[J]. Chinese Journal of Image and Graphics, 2019, 24(12):2255-2266.
[4]	朱明, 焦会敏, 赵兴运. 基于邻域优先搜索和纹理相似度匹配的图像颜色转移方法[J]. 影像科学与光化学, 2020, 38(6):935-940.
	ZHU Ming, JIAO Huimin, ZHAO Xingyun. A color transfer method based on neighborhood-first searching and texture similarity matching[J]. Imaging Science and Photochemistry, 2020, 38(6):935-940.
[5]	赵浩如, 张永, 刘国柱. 基于RPN与B-CNN的细粒度图像分类算法研究[J]. 计算机应用与软件, 2019, 36(3):216-219.
	ZHAO Haoru, ZHANG Yong, LIU Guozhu. Research on fine-grained image classification algorithm based on RPN and B-CNN[J]. Computer Applications and Software, 2019, 36(3):216-219.
[6]	王安琪. 基于图像内容的服装分类和推荐方法研究[D]. 昆明:昆明理工大学, 2017: 19.
	WANG Anqi. Research on clothing classification and recommendation method based on image content[D]. Kunming: Kunming University of Science and Technology, 2017: 19.
[7]	刘晓刚. 基于服装品牌的设计元素理论研究[D]. 上海:东华大学, 2005:76-80.
	LIU Xiaogang. The research on design elements theory based on fashion brands[D]. Shanghai: Donghua University, 2005:76-80.
[8]	谢珍珍, 张颖, 张思潮. 服装风格量化研究综述[J]. 现代装饰(理论), 2014(11):232-233.
	XIE Zhenzhen, ZHANG Ying, ZHANG Sichao. Summary of quantitative research on clothing style[J]. Modern Decoration (Theory), 2014(11):232-233.
[9]	冯利, 刘晓刚. 服装风格的量化方法初探[J]. 东华大学学报(自然科学版), 2004(1):57-61.
	FENG Li, LIU Xiaogang. Methods of measure costume style[J]. Journal of Donghua University(Natural Science), 2004(1):57-61.
[10]	贾玺增, 陈建辉. 服装风格与面料特征[J]. 天津工业大学学报, 2002(5):63-65.
	JIA Xizeng, CHEN Jianhui. Dress style and dress material character[J]. Journal of Tianjin Polytechnic University, 2002(5):63-65.
[11]	陈雁, 李栋高. 服装颜色的感觉生理研究[J]. 纺织学报, 2004, 25(3):68-69.
	CHEN Yan, LI Donggao. Study on the sensory physiology of clothing color[J]. Journal of Textile Research, 2004, 25(3):68-69.
[12]	吴圣美, 刘骊, 付晓东, 等. 结合人体检测和多任务学习的少数民族服装识别[J]. 中国图象图形学报, 2019, 24(4):562-572.
	WU Shengmei, LIU Li, FU Xiaodong, et al. Human detection and multi-task learning for minority clothing recognition[J]. Journal of Image and Graphics, 2019, 24(4):562-572.
[13]	张凯丽. 基于深度学习的服装属性识别与关键点定位算法的研究[D]. 杭州: 浙江工商大学, 2019:23-26.
	ZHANG Kaili. Research on clothing attribute recognition and landmark detection algorithm based on deep learning[D]. Hangzhou: Zhejiang Gongshang University, 2019: 23-26.
[14]	LIU Ziwei. DeepFashion:powering robust clothes recognition and retrieval with rich annotations [C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos: IEEE Computer Society, 2016:1096-1104.
[15]	林城龙, 胡伟, 李瑞瑞. 基于深度卷积神经网络的层次多任务服装分类[J]. 中国体视学与图像分析, 2018, 23(2):159-165.
	LIN Chenglong, HU Wei, LI Ruirui. Hierarchical multi-task clothing classification based on deep convolutional neural network[J]. Chinese Journal of Stereology and Image Analysis, 2018, 23(2):159-165.
[16]	夏明, 宋婧, 姜朝阳, 等. 基于连衣裙结构特征匹配的款式识别技术[J]. 纺织学报, 2020, 41(7):141-146.
	XIA Ming, SONG Jing, JIANG Zhaoyang, et al. Style recognition technique based on feature matching in dress construction[J]. Journal of Textile Research, 2020, 41(7):141-146.
[17]	CHEN L. Research on clothing image classification by convolu-tional neural networks [C]//2018 11th International Congress on Image and Signal Processing. Netherlands: BioMedical Engineering and Informatics (CISP-BMEI), 2018:1-5.
[18]	HE K. Deep residual learning for image recogni-tion [C]// ZHANG X, REN S, SUN J, et al. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Los Alamitos: IEEE Computer Society, 2016: 770-778.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

属性类别	训练集或测试集	准确率	平均精度	交叉熵损失率
Collar_design	训练集	0.876	0.931	0.350
Collar_design	测试集	0.805	0.890	0.533
Lapel_design	训练集	0.900	0.945	0.277
Lapel_design	测试集	0.844	0.910	0.494
Neck_design	训练集	0.835	0.907	0.458
Neck_design	测试集	0.728	0.843	0.700
Neck_line	训练集	0.831	0.902	0.484
Neck_line	测试集	0.808	0.886	0.594
Skirt_length	训练集	0.823	0.900	0.487
Skirt_length	测试集	0.768	0.867	0.646
Pant_length	训练集	0.874	0.932	0.355
Pant_length	测试集	0.810	0.897	0.533
Coat_length	训练集	0.793	0.882	0.592
Coat_length	测试集	0.770	0.870	0.652
Sleeve_length	训练集	0.841	0.912	0.449
Sleeve_length	测试集	0.834	0.909	0.456

对比图像	相似度计算方法	相似度
图7(a)、(b)	基于本文的款式风格特征	0.933
图7(a)、(b)	传统的图像余弦相似度	0.881
图7(c)、(d)	基于本文的款式风格特征	0.684
图7(c)、(d)	传统的图像余弦相似度	0.906