The Application of Deep and Transfer Learning for Identifying Cashmere and Wool Fibers

ABSTRACT

Identification of wool and cashmere fibers is one of the most important topics in the textile industry. In order to recognize these similar fibers, a novel identification method based on the convolution neural network and deep learning was proposed in this paper. As we all know, training a new identification network commonly requires lots of sample images and needs an amount of time, so the transfer learning was adopted for the fiber identification. The four pre-trained convolution neural networks, which consist of AlexNet, VGG-16, VGG-19, GoogLeNet, were used for the transfer learning to identify these similar fibers. Then, 65 fiber images of four kinds of fiber samples including goat hair, yellow wool, sheep wool, and cashmere fibers, were collected, respectively, and processed by the methods including random interception and rotation to obtain a total of 390 fiber images, respectively, for the experiment analysis. After comparing different network models, the results showed that the highest identification accuracy was 99.15%, obtained by the VGG-16 transfer learning model and the proportion of training set to testing set was 7:3. In addition, compared with the traditional machine learning algorithmics, this method also had a great improvement in the model performance and identification accuracy.

摘要

摘要羊毛和羊绒纤维的鉴别是纺织工业的重要课题之一. 为了识别这些相似的纤维，提出了一种基于卷积神经网络和深度学习的识别方法. 众所周知，训练一个新的识别网络通常需要大量的样本图像，并且需要花费大量的时间，因此采用传递学习进行纤维识别. 利用AlexNet、VGG-16、VGG-19、GoogLeNet四个预训练卷积神经网络进行传递学习，识别出这些相似的纤维. 然后分别采集山羊毛、黄毛、绵羊毛和羊绒4种纤维样品的65幅纤维图像，采用随机截取和旋转等方法进行处理，得到390幅纤维图像，进行实验分析. 通过对不同网络模型的比较，结果表明，VGG-16迁移学习模型的识别率最高为99.15%，训练集与测试集的比例为7:3. 此外，与传统的机器学习算法相比，该方法在模型性能和辨识精度上也有很大的提高.

KEYWORDS:

• Cashmere/wool identification
• deep learning
• transfer learning
• convolution neural networks

关键词:

• 羊绒/羊毛识别
• 深度学习
• 迁移学习
• 卷积神经网络

Acknowledgments

This work was supported by the National Natural Science Foundation of China, under Grant 61876106, the Shanghai Natural Science Foundation of China, under Grant 18ZR1416600, the Zhihong Scholars Plan of Shanghai University of Engineering Science, under Grant 2018RC032017, and the Shanghai Local capacity-building projects, under Grant 19030501200.