纺织学报 ›› 2023, Vol. 44 ›› Issue (08): 110-117.doi: 10.13475/j.fzxb.20220400301

• 纺织工程 • 上一篇    下一篇

基于SURF和VLAD特征编码的面料图案检索研究

赵文浩, 向军, 张宁, 潘如如()   

  1. 江南大学 纺织科学与工程学院, 江苏 无锡 214122
  • 收稿日期:2022-04-01 修回日期:2023-05-06 出版日期:2023-08-15 发布日期:2023-09-21
  • 通讯作者: 潘如如(1982—),男,教授,博士。主要研究方向为数字化纺织技术。E-mail: prrsw@163.com。
  • 作者简介:赵文浩(1997—),男,硕士。主要研究方向为面料图像检索。
  • 基金资助:
    国家自然科学基金项目(61976105);中国纺织工业联合会应用基础研究项目(J202006)

Research on fabric pattern retrieval based on SURF and VLAD feature coding

ZHAO Wenhao, XIANG Jun, ZHANG Ning, PAN Ruru()   

  1. College of Textile Science and Engineering, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Received:2022-04-01 Revised:2023-05-06 Published:2023-08-15 Online:2023-09-21

RichHTML

3

PDF (PC)

29

摘要：

为解决纺织面料图案检索算法检索效率低下、精度低等问题,提出一种基于SURF和VLAD特征编码的纺织面料图案检索算法。首先构建带有图案的面料数据库,并提取图像的SURF特征以对图像内容进行表达;接着对采集的原始面料进行聚类生成视觉词典,由生成的视觉词典对数据库中面料图像的SURF特征进行VLAD特征编码,以聚合生成VLAD向量;然后在保证VLAD对图像表达能力的前提下,对生成的VLAD向量进行主成分分析以降低向量维度,提高检索效率;最后采用Ball-tree算法构建索引,加快匹配速度,提高检索效率。实验结果显示,在视觉词典规模为512,保留维度数为512时,算法平均检索精度达到了83.5%,平均检索时间为0.488 s。

关键词: 面料检索, SURF特征, VLAD特征编码, 面料图案, 主成分分析

Abstract:

Objective With the increasing proportion of patterned fabrics produced in fabric manufacturing enterprises, there have been more requirements relating to the production and management of such fabrics, including querying the production materials of imitation, reducing the inventory management costs. The current textile pattern retrieval algorithm has low retrieval efficiency and poor adaptability, which is not conducive to the retrieval of fabric patterns by fabric enterprises. Therefore, a fast and accurate retrieval algorithm for similar patterns was proposed to help enterprises improve the utilization rate of production materials and reduce the cost of fabric inventory management, and promote textile fabric enterprises to enhance the competitiveness of the industry.

Method This research proposed a textile pattern retrieval algorithm based on SURF (speeded up robust features) and VLAD (vector of locally aggregated descriptors) feature coding. A fabric database with patterns was constructed, and SURF features of the image were extracted to express the image content, before the original fabric was clustered to generate a visual dictionary. The generated visual dictionary encoded the SURF features of the fabric image in the database with VLAD features to aggregate to generate VLAD vectors. On the premise of ensuring the ability of VLAD to express images, principal component analysis was performed on the generated VLAD vectors to reduce the vector dimensions and improve the retrieval efficiency. Finally, the Ball tree algorithm was used to build the index to accelerate the matching speed and improve the retrieval efficiency.

Results In order to obtain good retrieval performance, the parameters of the algorithm used were optimized. Through the comparison of retrieval precision and time, it was seen that within a certain range the precision of fabrics was improved with the increase of dictionary size K. However, when K was increased to 1 024, the precision only increased slightly with the retrieval time doubled. As the size of the visual dictionary was increased, the complexity and time cost of training the visual dictionary were greatly increased. Considering comprehensively, the dictionary size K=512 was taken in this research. From the perspectives of the retrieval precision and time, it was seen that with the reduction of feature dimensions, the precision of fabrics was constantly decreasing. When the dimension was reduced to 64, the precision of fabrics decreased greatly, with a slow reduction in retrieval time. Considering the effectiveness and time complexity of the algorithm, the feature dimension D=512 was selected in the subsequent experiments of this research. In order to highlight the advantages of the algorithm in this paper, this paper compared the "ORB (Oriented FAST and Rotated BRIEF)+VLAD", "HOG (Histogram of Oriented Gradient)+VLAD", "Color Moments (CM)", "Color Moments+ Gray-level Co-occurrence Matrix" and "Color Histogram+ Gray-level Co-occurrence Matrix" for retrieval verification. The average retrieval accuracy of the algorithm reached 83.5%, the average retrieval time was 0.488s, and the comprehensive performance was better than other algorithms. From the P-R curves, it was evident that the algorithm developed in this research performed the best on the data set built. "HOG+VLAD" was slightly better than "ORB+VLAD", indicating that SURF was able to better describe the characteristics of patterned fabrics than ORB and HOG.

Conclusion The experimental results show that when the scale of visual dictionary K is 512 and the number of reserved dimensions is 512, the average retrieval accuracy of the algorithm for fabric pattern retrieval is 83.5%, the average retrieval time is 0.488 seconds, and the algorithm has good size invariance and rotation invariance. It has good performance and adaptability for fabric pattern retrieval, which has certain practical significance for the practical application of textile enterprises.

Key words: fabric retrieval, SURF feature, VLAD feature coding, fabric pattern, principal component analysis

中图分类号: 

  • TS941.26

图1

尺度空间的构建与特征点的定位"

图2

特征描述子的生成"

图3

码本的生成"

图4

VLAD特征算法的生成"

图5

数据集的球树结构"

图6

图案类面料检索算法流程"

表1

不同视觉词典规模下的查准率和时间对比"

词典规模(K) 查准率/% 平均检索时间/s
32 86.3 0.382
64 90.5 0.401
128 93.8 0.413
256 94.3 0.434
512 95.5 0.488
1 024 95.9 0.862

表2

维度对查准率和时间的影响"

维度(D) 查准率/% 平均检索时间/s
1 024 95.8 0.753
512 95.5 0.488
256 92.9 0.407
128 88.3 0.378
64 81.6 0.369

图7

检索效果展示 注:数字1~10代表检索结果返回顺序。"

表3

不同算法的平均检索精度和检索时间对比"

算法 平均检索精度/% 检索时间/s
本文算法 83.5 0.488
ORB+VLAD 66.8 0.358
HOG+VLAD 70.2 0.523
CM 40.3 0.263
文献[14]方法 56.7 0.627
文献[15]方法 63.8 0.755

图8

不同算法的检索效果 注:a—待检图像;b—本文算法检索结果;c—文献[14]算法检索结果;d—文献[15]算法检索结果。 数字1~10为检索结果返回顺序。"

图9

不同算法的P-R曲线图"

[1] SZTANDERA L M, CARDELLO A V, WINTERHALTER C, et al. Identification of the most signifycant comfort factors for textiles from processing mechanical, handfeel, fabric construction, and perceived tactile comfort data[J]. Textile Research Journal, 2012, 83(1): 34-43.
doi: 10.1177/0040517512438121
[2] XIANG J, ZHANG N, PAN R, et al. Fabric image retrieval system using hierarchical search based on deep convolutional neural network[J]. IEEE Access, 2019, 7: 35405-35417.
doi: 10.1109/Access.6287639
[3] 张宁. 基于特征工程的机织面料图像检索方法研究[D]. 无锡: 江南大学, 2021:1-50.
ZHANG Ning. Research on image retrieval method of woven fabric based on Feature Engineering[D]. Wuxi: Jiangnan University, 2021:1-50.
[4] 康锋, 胡洁, 张华熊, 等. 基于最大稳定极值区域的织物图案检索[J]. 纺织学报, 2015, 36(10): 156-160.
KANG Feng, HU Jie, ZHANG Huaxiong, et al. Fabric pattern retrieval based on maximum stable extremum region[J]. Journal of Textile Research, 2015, 36(10): 156-160.
[5] 向忠, 何旋, 钱淼, 等. 基于边缘和颜色特征的织物印花花型检索[J]. 纺织学报, 2018, 39(5): 137-143.
XIANG Zhong, HE Xuan, QIAN Miao, et al. Fabric printing pattern retrieval based on edge and color features[J]. Journal of Textile Research, 2018, 39(5): 137-143.
[6] 曹霞, 李岳阳, 罗海驰, 等. 蕾丝花边的改进型纹理特征检索方法[J]. 纺织学报, 2016, 37(6): 142-147,154.
CAO Xia, LI Yueyang, LUO Haichi, et al. Improved texture feature retrieval method for lace[J]. Journal of Textile Research, 2016, 37(6): 142-147,154.
[7] 刘颖, 倪天宇, 王富平, 等. 融合局部聚合描述符和全局特征的现勘图像分类算法[J]. 科学技术与工程, 2020, 20(3): 1118-1124.
LIU Ying, NI Tianyu, WANG Fuping, et al. An image classification algorithm based on local aggregation descriptors and global features[J]. Science Technology and Engineering, 2020, 20(3): 1118-1124.
[8] 江南大学. 基于SURF和VLAD特征编码的面料图案检索算法: CN202210889749. 0[P]. 2022-11-01.
Jiangnan University. Fabric pattern retrieval algorithm based on SURF and VLAD feature coding: CN202210889749. 0[P]. 2022-11-01.
[9] BAY H, TUYTELAARS T, LUC Van Gool. SURF: speeded up robust features[J]. Computer Vision and Image Understanding, 2008, 110(3): 346-359.
doi: 10.1016/j.cviu.2007.09.014
[10] 修瑛昌, 杨文静. Mini Batch K-means算法在遥感影像分类中的应用[J]. 鲁东大学学报(自然科学版), 2017, 33(4): 359-363.
XIU Yingcang, YANG Wenjing. Application of mini batch k-means algorithm in remote sensing image classification[J]. Journal of Ludong University, 2017, 33(4): 359-363.
[11] JÉGOU H, DOUZE M, SCHMID C, et al. Aggregating local descriptors into a compact image representa-tion[C]// 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, CA:IEEE. 2010: 3304-3311.
[12] RETONDARO L C d S C, ESPERANÇA C. Optimized 2D Ball Trees[J]. Computer Science, 2021. 10.1109/sibgrapi54419.2021.00014.
[13] 房梦玉, 马明栋. 改进的PCA-LDA人脸识别算法的研究[J]. 计算机技术与发展, 2021, 31(2): 65-69.
FANG Mengyu, MA Mingdong. Research on improved PCA-LDA face recognition algorithm[J]. Computer technology and development, 2021, 31 (2): 65-69.
[14] AHSANI A F, SARI Y A, ADIKARA P P. Food image retrieval with gray level co-occurrence matrix texture feature and CIE L*a*b* color moments feature[C]//2019 International Conference on Sustainable Information Engineering and Technology (SIET). Malang:IEEE, 2019: 130-134.
[15] YUE J, LI Z, LIU L, et al. Content-based image retrieval using color and texture fused features[J]. Mathematical and Computer Modeling, 2011, 54: 1121-1127.
doi: 10.1016/j.mcm.2010.11.044
[1] 张晓侠, 刘凤坤, 买巍, 马崇启. 基于BP神经网络及其改进算法的织机效率预测[J]. 纺织学报, 2020, 41(08): 121-127.
[2] 万莉, 贡丽英, 贾敏瑞. 基于主成分分析的智能复合材料结构损伤类型识别[J]. 纺织学报, 2019, 40(05): 53-58.
[3] 张陆佳 林兰天 陈春敏 申炎仃 高琮. 基于主成分分析的纤维拉伸断裂声发射信号识别[J]. 纺织学报, 2018, 39(01): 19-24.
[4] 夏凤勤 毋戈 谢昊洋 钟跃崎. 基于人体纵截面特征曲线的体型分类[J]. 纺织学报, 2017, 38(06): 86-91.
[5] 万振凯 贡丽英 万莉. 基于损伤指数的三维编织复合材料结构损伤评估[J]. 纺织学报, 2017, 38(05): 69-74.
[6] 李东 万贤福 汪军. 采用傅里叶描述子和支持向量机的服装款式识别方法[J]. 纺织学报, 2017, 38(05): 122-127.
[7] 王玲 杨以雄 陈炜. 服装工序相似性标准工时预测[J]. 纺织学报, 2016, 37(11): 114-119.
[8] 苟捷 刘建立 高卫东. 应用主成分分析的原棉可纺性指数构建[J]. 纺织学报, 2015, 36(08): 16-21.
[9] 金关秀 应伟伟 雷新 从明芳 祝成炎. 纺粘非织造布孔隙形状的定量分析[J]. 纺织学报, 2015, 36(08): 56-61.
[10] 夏明 王文玲 张文斌. 人体特征断面形状拟合及围度尺寸预测[J]. 纺织学报, 2014, 35(8): 76-0.
[11] 钟智丽 王玉新. 聚丙烯长丝/芳纶包缠纱捻度的优化[J]. 纺织学报, 2014, 35(6): 40-0.
[12] 耿艳蕾;崔运花;程隆棣;黄淑韵. 不同结构竹浆纤维纱线性能测试及分类[J]. 纺织学报, 2010, 31(4): 35-38.
[13] 刘贵;于伟东.. 主成分分析的BP神经网络在精毛纺粗纱工序中的应用[J]. 纺织学报, 2008, 29(9): 34-37.
[14] 徐松杰. 基于主成分分析法的纺织企业技术改造效果评价[J]. 纺织学报, 2008, 29(7): 128-131.
[15] 李艳梅.;张渭源. 基于灰色模型的服装面料缝纫皱缩预测[J]. 纺织学报, 2008, 29(6): 92-95.
Viewed
Full text


Abstract

Cited
  Shared   
  Discussed   
No Suggested Reading articles found!
京ICP备14006648号　 京公网安备11010502044800号
版权所有 © 2021 《纺织学报》编辑部
地址： 北京市朝阳区延静里中街3号主楼6层《纺织学报》杂志社　邮政编码：100025　
电话：010-65017711,65917740　传真：010-65016539　Email:fangzhixuebao@vip.126.com
本系统由北京玛格泰克科技发展有限公司设计开发