摘要
在样本数据稀缺的情况下,为快速准确地获取古代汉服的关键尺寸数据,提出一种基于卷积神经网络的汉服尺寸测量方案。首先搭建1个二阶段卷积神经网络模型GlobalNet-RefineNet进行关键点检测,通过2次迁移学习和反复迭代训练提高关键点识别准确率;再利用算法得到坐标点的像素距离,结合博物馆或发掘报告中给出的汉服平铺图和至少1个真实测量尺寸,通过比例映射得到全衣的尺寸数据。以汉服上衣为例进行实验验证,结果表明:经过2次迁移学习,卷积神经网络模型的收敛程度高,训练效果好,通过该方案测得的汉服上衣尺寸相对误差在0.59%~4.17%之间;该方案为传统服饰的复原研究和文物尺寸测量工作提供了新思路。
In order to quickly and accurately obtain the key dimensions of the ancient Chinese Han-style costumes with scarce sample data,a clothing size measurement scheme based on the use of convolutional neural network was proposed in this paper.Firstly,a two-stage convolutional neural network model GlobalNet-RefineNet was built for detecting the key points.The accuracy of the key point recognition was improved through twice transfer learning and repeated iterative training.An algorithm was used to get the pixel distance between coordinate points.Combined with the tiles of Han-style costume and at least one real measurement size given in the museum or excavation report,the size data of the whole garment were obtained through proportional mapping.This research used the top of a Han-style costume as an example for experiments.The research results show that after two times of transfer learning,the model has a high degree of convergence and good training effect.The relative error of costume top size measured by this scheme is between 0.59%-4.17%.This research provides new ideas for the restoration research of traditional clothing and the measurement of cultural relics.
作者
王奕文
罗戎蕾
康宇哲
WANG Yiwen;LUO Ronglei;KANG Yuzhe(School of Fashion Design & Engineering, Zhejiang Sci-Tech University, Hangzhou, Zhejiang 310018, China;School of International Education, Zhejiang Sci-Tech University, Hangzhou, Zhejiang 310018, China;Silk and Fashion Culture Center, Hangzhou, Zhejiang 310018, China;School of Information Science and Technology, Zhejiang Sci-Tech University, Hangzhou, Zhejiang 310018, China)
出处
《纺织学报》
EI
CAS
CSCD
北大核心
2020年第12期124-129,共6页
Journal of Textile Research
基金
浙江省“十三五”高校虚拟仿真实验教学项目(浙教办函[2019]365号)。
关键词
尺寸测量
服装关键尺寸
汉服
卷积神经网络
迁移学习
dimensional measurement
key dimensions of costume
Han-style costume
convolutional neural network
transfer learning
