|
|
| |
|
|
| 本文已被：浏览 337次 下载 425次 |
码上扫一扫！ |
|
|
| 基于改进Cascade R-CNN的织物瑕疵检测方法 |
|
张银, 佟乐
|
|
上海师范大学 信息与机电工程学院, 上海 201418
|
|
| 摘要: |
| 制造业瑕疵检测问题是工业产品的关键环节，其中织物瑕疵检测尤为关键.针对织物图像尺寸大、局部瑕疵数量少、图像背景复杂等问题，设计了一种基于Cascade R-CNN融合尺度不变特征的织物瑕疵检测方法.通过训练具有多个递增交并比（IoU）阈值的级联检测网络，解决了传统深度神经网络的过拟合和错配问题；针对工业生产上存在的待检测图片发生倾斜、偏转和迁移等问题，通过尺度不变特征变换（SIFT）算法对待检测织物图进行预处理，将其正确定位到模板图上对应位置后，再进行瑕疵检测；借鉴孪生学习的思路，通过扩充输入通道，同时输入并分析待检测的图像和模板，提高了瑕疵检测的精度.实验结果表明：该模型能够有效地完成瑕疵检测任务，可对7类织物瑕疵进行检测与分类，平均精度均值达87.5%. |
| 关键词: 织物瑕疵检测 Cascade R-CNN 尺度不变特征变换(SIFT) 孪生学习 |
| DOI：10.3969/J.ISSN.1000-5137.2023.02.013 |
| 分类号:TS101.97;TP183 |
| 基金项目:上海市青年科技英才扬帆计划（19YF1437100） |
|
| Fabric defect detection based on improved Cascade R-CNN |
|
ZHANG Yin, TONG Le
|
|
College of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Shanghai 201418, China
|
| Abstract: |
| Manufacturing defect detection was a key part of industrial products, among which fabric defect detection was particularly critical. Aiming at the problems of large fabric image size, small number of local defects and complex image background, a fabric defect detection method based on Cascade R-CNN fusion scale invariant feature transform(SIFT) was designed. The overfitting and mismatch problems of traditional deep neural networks could be solved by Cascade R-CNN through training cascading detection networks with multiple increasing intersection over union (IoU) thresholds. In view of the tilt, deflection and migration of the picture to be detected in industrial production, the SIFT algorithm was used to preprocess the fabric diagram to be detected, which was correctly positioned in the corresponding points on the template diagram and then detected. According to the idea of twin learning, the accuracy of defect detection was improved by expanding the input channel and simultaneously entering the image and template to be detected for analysis. Experimental results showed that the model could effectively complete the defect detection task, which could detect and classify 7 types of fabric defects, with an average accuracy of 87.5%. |
| Key words: fabric defect detection Cascade R-CNN scale invariant feature transform(SIFT) twin learning |
|
Home