一种端到端的织物颜色自动预测框架  

An end-to-end automatic fabric color prediction framework

作　　者：周奕军 陈昭 李悦[1] 纪柏林 王国栋 刘国华[1] ZHOU Yijun;CHEN Zhao;LI Yue;JI Bolin;WANG Guodong;LIU Guohua(School of Computer Science and Technology,Donghua University,Shanghai 201620,China;College of Chemistry,Chemical Engineering and Biotechnology,Donghua University,Shanghai 201620,China)

机构地区：[1]东华大学计算机科学与技术学院,上海201620 [2]东华大学化学化工与生物工程学院,上海201620

出　　处：《东华大学学报（自然科学版）》2021年第2期51-57,64,共8页Journal of Donghua University(Natural Science)

基　　金：国家重点研发计划资助项目(2017YFB0309800);国家自然科学基金资助项目(61702094);上海市科学技术委员会科研计划资助项目(17YF1427400)。

摘　　要：鉴于传统的织物染料配方设计流程耗时耗力,提出一种端到端的机器学习框架,从染色工艺参数自动预测产品颜色出发,为染料配方工作提供精准的理论指导。理论上,该框架可由任意有监督回归模型实现。采用多维支持向量回归机(multi-dimensional support vector regressor,M-SVR)或反向传播神经网络(back propagation neural network,BPNN)对织物的颜色进行回归分析和预测,并提出用遗传算法(genetic algorithm,GA)优化神经网络参数初始值,从而提高网络的优化效率。从山东省华纺股份有限公司提供的真实生产资料中选取8种关键工艺参数和两种光照情形下的织物颜色CIELAB值,对模型进行训练和性能评估。结果表明:该框架最低预测误差可达0.48;在相同训练条件下,GA初始化的神经网络比随机初始化的网络耗时更少。由此可见,该框架有助于降低染料配方的误差,提高工作效率。In view of the traditional time-consuming and labor-intensive textile dye formulation process,an end-to-end machine learning framework is proposed to provide accurate theoretical guidance for the formulation process,starting from the automatic prediction of product color by dyeing process parameters.Theoretically,the framework can be embedded with an arbitrary supervised regression model.The multi-dimensional support vector regressor(M-SVR)and back propagation neural network(BPNN)were used for fabric color regression and prediction.A genetic algorithm(GA)was proposed to optimize the initial values of neural network parameters to improve the optimization efficiency of the network.The model was trained and its performance was evaluated by selecting eight key process parameters and CIELAB values for fabric color under two lighting conditions from real production data provided by Shandong Huafang Co.Ltd.The results show that the framework has a minimum prediction error of 0.48;under the same training conditions,it takes less time for the GA to initialize a neural network than a randomly initialized network.Thus,the framework is effective in reducing dyeing recipe errors and improving work efficiency.

关 键 词：织物颜色预测 端到端 机器学习 负反馈神经网络 多维支持向量回归机 遗传算法 

分 类 号：TS193.5[轻工技术与工程—纺织化学与染整工程]