摘要
为了解决传统方法在新疆游牧民族纺织品纹样分类时准确率低、速度慢的问题,提出一种改进的卷积神经网络模型(ResNet 18-CA)。在ResNet 18卷积神经网络模型基础上添加注意力机制模块,加强对纹样特征的提取,并引入迁移学习思想,有效防止网络过拟合,并将改进后的网络模型与当前经典的几类卷积神经网络模型进行对比实验。改进后的卷积神经网络模型在建立的新疆传统纺织品纹样数据集上的分类识别准确率达到了98.62%,相比于原始的ResNet 18模型提高了3.72%,而模型大小仅增大0.2 MB,改进后的神经网络模型的分类准确率更高。
To address the issues of low accuracy and slow speed in traditional methods for pattern recognition and classification of textile patterns among nomadic ethnic groups in Xinjiang,an improved convolutional neural network(ResNet 18-CA)for the recognition of textile patterns among Xinjiang nomadic ethnic groups was proposed in this paper,which enhances the extraction of pattern features by adding an attention mechanism module to the ResNet 18 convolutional neural network model.Additionally,the concept of transfer learning to effectively prevent overfitting of the network was introduced in the paper.The improved convolutional neural network model has a classification recognition accuracy of 98.62% on the established Xinjiang traditional textile pattern data set,which is 3.72% higher than the original ResNet 18 model,but the model is only 0.2 MB larger,and the improved neural network model has a higher classification accuracy.
作者
赵楷文
薄贤姝
钱娟
阎明星
ZHAO Kaiwen;BO Xianshu;QIAN Juan;YAN Mingxing(School of Textiles and Clothing,Xinjiang University,Urumqi,Xinjiang 830046,China;Xinjiang Silk Road Costume Art Research Base,Xinjiang University,Urumqi,Xinjiang 830046,China)
出处
《毛纺科技》
CAS
北大核心
2024年第11期111-118,共8页
Wool Textile Journal
基金
国家社科基金项目(22BMZ164)
新疆大学中亚研究院项目(24FPY004)
大学生创新训练项目(XJU-SRT-23039)。
