基于卷积神经网络和迁移学习的色织物疵点检测  

Yarn-dyed fabric defect detection based on convolution neural network and transfer learning

作　　者：罗俊丽[1] 路凯[1] LUO Junli;LU Kai(College of Information Engineering,Xuchang University,Xuchang 461000,China)

机构地区：[1]许昌学院信息工程学院

出　　处：《上海纺织科技》2019年第6期52-56,共5页Shanghai Textile Science & Technology

基　　金：河南省高等学校重点科研项目(18A520010);河南省重点研发与推广专项(科技攻关)项目(192102210135)

摘　　要：针对色织物疵点检测准确率较低的问题,提出一种基于卷积神经网络的疵点识别方法。首先对织物图像进行预处理,然后将无疵点织物和5类疵点织物图像样本输入到残差网络模型进行训练和测试,最终总识别率为91.53%。以残差网络模型为例,研究了迁移学习在织物疵点识别中的应用。将ImageNet数据集上预训练的图像特征迁移到织物疵点识别任务上,并在两个大小不同的数据集上比较迁移学习的效果。结果显示,当数据集较小时,通过迁移学习可以提升模型的识别率。Aiming at the problem of low accuracy of yarn-dyed fabric defect detection, a method based on convolutional neural network is proposed. First, fabric images are preprocessed. Non-defective fabric samples and five kinds of defective fabric samples in the dataset are fed into the ResNet model, and the total recognition rate is 91.53%. Taking ResNet model as an example, the application of transfer learning in fabric defect detection is studied. The image features pre-trained on ImageNet dataset are transferred to ResNet-34 model for the task of fabric defect detection. The effects of transfer learning are evaluated on two datasets of different sizes. Experiments show that transfer learning could improve the performance of the model when the dataset is small.

关 键 词：色织物 织物疵点 检测 卷积神经网络 残差网络 迁移学习 

分 类 号：TS101.97[轻工技术与工程—纺织工程]