基于卷积神经网络的数码印花缺陷分类算法 下载: 858次
Digital Printing Defect Classification Algorithm Based on Convolutional Neural Network
西安工程大学电子信息学院, 陕西 西安 710048
图 & 表
图 1. 数码印花缺陷样例。(a) PASS道;(b)喷墨不均;(c)漏墨;(d)布匹褶皱
Fig. 1. Examples of digital printing defects. (a) PASS tracks; (b) uneven inkjet; (c) ink leakage; (d) fabric wrinkles
下载图片 查看原文
图 2. RGB颜色空间直方图均衡化处理结果。(a) PASS道;(b)喷墨不均;(c)漏墨;(d)布匹褶皱
Fig. 2. RGB color space histogram equalization processing results. (a) PASS tracks; (b) uneven inkjet; (c) ink leakage; (d) fabric wrinkles
下载图片 查看原文
图 3. 高斯滤波处理结果。(a) PASS道;(b)喷墨不均;(c)漏墨;(d)布匹褶皱
Fig. 3. Gaussian filtering processing results. (a) PASS tracks; (b) uneven inkjet; (c) ink leakage; (d) fabric wrinkles
下载图片 查看原文
图 4. 局部均值法的图像分辨率调整结果。(a)分辨率调整前;(b)分辨率调整后
Fig. 4. Adjustment results of image resolution based on local mean algorithm. (a) Before resolution adjustment; (b) after resolution adjustment
下载图片 查看原文
图 5. 图像数据增强结果。(a)原图;(b)垂直翻转;(c)水平镜像;(d)旋转90°;(e)旋转180°;(f)旋转270°
Fig. 5. Image data enhancement results. (a) Original image; (b) flip vertically; (c) horizontal mirroring; (d) rotate 90°; (e) rotate 180°; (f) rotate 270°
下载图片 查看原文
图 6. 分类算法流程图
Fig. 6. Flow chart of classification algorithm
下载图片 查看原文
图 7. 卷积神经网络拓扑结构
Fig. 7. Topological structure of convolutional neural network
下载图片 查看原文
图 8. 数码印花缺陷数据集样本。(a)~(d) PASS道;(e)~(h)喷墨不均;(i)~(l)漏墨;(m)~(p)布匹褶皱
Fig. 8. Samples of digital printing defect data set. (a)--(d) PASS tracks; (e)--(h) uneven inkjet; (i)--(l) ink leakage; (m)--(p) fabric wrinkles
下载图片 查看原文
图 9. 总损失率曲线
Fig. 9. Total loss rate curve
下载图片 查看原文
图 10. 不同CNN模型预测Kappa系数值
Fig. 10. Kappa coefficient value predicted by different CNN models
下载图片 查看原文
表 1数码印花缺陷特征对比
Table1. Comparison of defect features in digital printing
Type of defect | Cause of formation | Appearance shape | Probability of occurrence |
---|
PASS tracks | Nozzle clogging,motor step deviation | Narrow linear | High | Uneven inkjet | Uneven inkjet output debugging | Flat | Low | Ink leakage | Inkjet pressure instability | Dotted | Medium | Fabric wrinkles | Uneven cloth press | Strip | Low |
|
查看原文
表 2不同目标函数对应的分类准确率
Table2. Classification accuracy corresponding to different objective functions
Objective function | Accuracy/% |
---|
Softmax cross entropy | 98.14 | Classification cross entropy | 96.42 | Binary cross entropy | 81.29 | Mean square loss | 88.02 | Hinge loss | 74.92 | ROC AUC score | 77.33 |
|
查看原文
表 3不同优化函数对应的分类准确率
Table3. Classification accuracy corresponding to different optimization algorithms
Optimization | Accuracy/% |
---|
Adaptive moment estimation | 98.21 | Stochastic gradient descent | 74.84 | Root mean square propagation | 65.38 | Momentum gradient descent | 92.73 | Adaptive sub-gradient method | 81.67 |
|
查看原文
表 4每类缺陷分类性能指标
Table4. Performance index of each defect classification
Defectclassification | Performance /% | Averageaccuracy /% | Standarddeviation |
---|
1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
---|
Validation set | 98.17 | 98.53 | 96.33 | 95.00 | 98.33 | 96.17 | 95.61 | 98.41 | 95.27 | 96.18 | 96.80 | 0.0133 | Test set | PASS tracks | 92 | 94 | 89 | 95 | 85 | 93 | 86 | 90 | 88 | 91 | 90.30 | 0.0316 | Uneven inkjet | 94 | 98 | 97 | 96 | 91 | 89 | 92 | 92 | 93 | 90 | 93.20 | 0.0286 | Ink leakage | 98 | 100 | 93 | 97 | 94 | 95 | 100 | 98 | 96 | 97 | 96.80 | 0.0223 | Fabric wrinkles | 100 | 93 | 96 | 95 | 98 | 96 | 97 | 94 | 95 | 94 | 95.80 | 0.0199 |
|
查看原文
表 5不同CNN模型训练和测试用时
Table5. Training and testing time of different CNN models
CNN model | LeNet5 | AlexNet | VGG16 | GoogLeNet | Proposed |
---|
Training/min | 76 | 92 | 114 | 136 | 65 | Testing/ms | 15 | 64 | 153 | 124 | 10 |
|
查看原文
苏泽斌, 高敏, 李鹏飞, 景军锋, 张缓缓. 基于卷积神经网络的数码印花缺陷分类算法[J]. 激光与光电子学进展, 2020, 57(24): 241011. Zebin Su, Min Gao, Pengfei Li, Junfeng Jing, Huanhuan Zhang. Digital Printing Defect Classification Algorithm Based on Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2020, 57(24): 241011.