基于改进Res-UNet网络的织物瑕疵图像识别方法
Image recognition method for fabrics defects based on improved Res-UNet network
于光许 1张富宇1
作者信息
- 1. 河南经贸职业学院,河南 郑州 450018
- 折叠
摘要
复杂花色织物的纹理和色彩常常是非规则的,导致织物表面瑕疵识别难度较高.针对上述问题,研究一种基于改进Res-UNet网络的织物表面瑕疵图像识别方法.采集织物图像并对其实施灰度化、去噪以及直方图均衡化处理,利用蝙蝠算法求取最佳提取网络层数,通过增加特征提取网络层数改进Res-UNet网络,利用改进后的Res-UNet网络识别织物表面瑕疵,并且采用迁移学习算法进一步优化识别模型的参数,实现织物表面瑕疵准确识别.结果表明:本文方法应用下,无论是素色样本,还是花色样本,其识别系数均达到 0.9 以上,相比基于标签嵌入方法的织物瑕疵识别方法和双路高分辨率转换网络的布匹瑕疵检测方法,本文方法对复杂花色样本的轮廓系数识别更高,适用性更好,识别能力更强.
Abstract
The texture and color of complex patterned fabrics are often irregular,making it difficult to identify surface defects.To address the above issues,a textile surface defect image recognition method based on an improved Res UNet network was studied.Textile images were collected and processed with gray-scale,de-noising and histogram equalization.The bat algorithm was used to obtain the optimal extraction network layer number.The Res-UNet network was improved by increasing the feature extraction network layer number,and the improved Res-UNet network was used to identify textile surface defects to achieve accurate identification of textile surface defects.The results show that under the application of the proposed method,the identification coefficient of both plain color samples and color samples is above 0.9.Compared with the textile defect identification method based on label embedding method and the fabric defect detection method based on dual-channel high-resolution conversion network,the identification method based on the improved Res-UNet network can identify the contour coefficient of complex color samples better.The applicability is better,the recognition ability is stronger.
关键词
改进Res-UNet网络/织物表面瑕疵/图像采集/预处理/图像识别Key words
improving Res-UNet network/fabrics surface defects/image acquisition/preprocessing/image recognition引用本文复制引用
出版年
2024
NETL
NSTL
万方数据