基于迁移学习的棉/毛纤维自动识别方法
Automatic identification method of cotton/wool fibers based on transfer learning
游小荣 1李淑芳2
作者信息
- 1. 常州纺织服装职业技术学院,江苏 常州 213164
- 2. 常州纺织服装职业技术学院,江苏 常州 213164;常州市生态纺织技术重点实验室,江苏 常州 213164
- 折叠
摘要
针对纺织品废料中织物纤维手工分类存在效率低、主观性强等问题,提出了一种小样本条件下纺织品废料中棉/毛纤维的自动识别方法.首先使用扫描电子显微镜,对已有的棉/毛纤维进行拍照,形成小样本棉/毛纤维图像集;然后分别加载 4 种经ImageNet数据集训练过的模型进行迁移学习,保留或部分微调模型的网络参数,并基于小样本图像集进行训练和验证,生成棉/毛纤维的分类模型;最后基于准确率、精确率和召回率评价指标,对各种分类模型进行对比测试,选出最优分类模型,实现棉/毛纤维的自动识别.经过实验发现ResNetXt50 模型在模型训练过程中取得了最高的精确率,其值为 97.33%.对测试集进行测试,结果显示通过微调后的 4 种分类模型中,ResNet50 和ResNetXt50 的测试准确率可达 99.537%,验证了方法的有效性.
Abstract
Aiming at the problems of low efficiency and strong subjectivity in manual classification of textile fibers,an automatic identification method for cotton and wool fibers in textile wastes under the condition of small samples was proposed.Firstly,scanning electron microscope was used to photograph the existing cotton fiber and form a small sample of cotton fiber image set.Then,four kinds of models trained by ImageNet data set were loaded for transfer learning,network parameters of the models were retained or partially fine-tuned,and the classification model was generated based on the small sample image set for training and verification.Finally,based on the evaluation indexes of accuracy,accuracy and recall rate,various classification models were compared and tested,and the optimal classification model was selected to realize the automatic recognition of cotton fiber.The experimental results show that the ResNetXt50 model achieves the highest accuracy rate in the process of model training,and its value is 97.33%.The test set is tested,and the results show that the test accuracy of ResNet50 and ResNetXt50 can reach 99.537%among the four kinds of classification models after fine-tuning,which verifies the effectiveness of the method.
关键词
纺织品废料/小样本/迁移学习/预训练模型/图像识别Key words
textile waste/small sample/transfer learning/pretrained model/image identification引用本文复制引用
基金项目
江苏省高等学校大学生实践创新训练项目(202112807013Y)
江苏省碳纤维先进材料智能制造工程技术研究开发中心项目(苏教科[2023]11号)
出版年
2024
NETL
NSTL
万方数据