Non-woven Fabric Real-time Defects Detection Method and Framework Design
-
摘要: 无纺布生产过程中产生的疵点会严重影响产品质量并限制生产效率. 提高疵点检测的自动化程度对于无纺布的生产效率和质量管控至关重要. 传统疵点检测方法难以应对纹理、疵点类型以及环境变化等问题, 限制了其应用范围. 近年来基于卷积神经网络的方法在疵点检测领域得到了广泛应用, 具有泛化性强、准确度高的特点. 但是在无纺布生产过程中, 布匹宽度大、速度快的特点会产生大量图像数据, 基于卷积神经网络的方法难以实现实时检测. 针对上述难题, 本文提出了一种基于最大稳定极值区域分析与卷积神经网络协同的疵点实时检测方法, 并设计了分布式计算处理架构应对数据流过大的问题. 在实际生产部署应用中, 本文所设计的系统与算法无需使用专用计算硬件(GPU、FPGA等), 通过8台工控机与16路工业摄像头对复卷机上布宽2.8 m、速度30 m/min的无纺布进行分布式实时在线检测, 大幅度提高无纺布生产中疵点检测的自动化程度与效率. 本文所提出的系统能够实现对0.3 mm以上疵点召回率100%, 对0.1 mm丝状疵点召回率98.8%.Abstract: The defects generated during the production process of non-woven fabric will seriously affect the quality and limit the efficiency. How to improve the automatic degree of non-woven fabric defects detection plays a significant role. The traditional defects detection methods cannot deal with the changing of texture, defects type and environments, which limits the application scope. In recent years, the methods based on Convolutional Neural Networks (CNN) have been widely used in the field of defects detection, which are shown to have the characteristics of strong generalization ability and high accuracy. However, in the non-woven fabric production process, the large width and high speed of cloth will introduce huge amount of image data, which makes it difficult for CNN based methods to achieve real-time detection. In this paper, a real-time defects detection method based on stable extremal region analysis and CNN is proposed, and a distributed computing architecture is designed to handle the problem of large image data stream. In the actual deployment application, the system designed in this paper does not need specific computing hardware (GPU, FPGA, etc.). 8 industrial computers and 16 industrial cameras are coupled together in a distribution scheme to finish real-time defects detection of non-woven fabric rewinder with cloth width 2.8 m and speed 30 m/min, which greatly improves the automation and production efficiency. The system proposed in this paper can achieve 100% recall rate of punctiform defects above 0.3 mm and 98.8% recall rate of 0.1 mm filiform defects.
-
图 3 本文采用的预检测模型结构图. 输入疵点图像经过三个不同尺度的卷积后得到三个特征图, 特征图拼接后作为Dense Block的输入. Dense Block输出与全连接层和softmax层相连. 其中虚线矩形框出部分为展开Dense Block的具体形式.
Fig. 3 The structure diagram of the pre-detection model used in this paper. The input defect image is convolved at three different scales to obtain three feature maps, and the feature maps are concatenated as the input of the Dense Block. The output of the Dense Block is connected to the fully connected layer and the softmax layer. Among them. The dashed rectangle outlines the specific form of the Dense Block.
图 5 NMS算法处理效果图. (a)为原图, (b)为网络预测效果图, 网络预测的矩形用红色矩形绘制出, (c)为经过NMS算法处理后的效果图.
Fig. 5 NMS algorithm processing effect diagram.(a) is the original image, (b) is the network prediction effect diagram, the network prediction rectangle is drawn with a red rectangle, and (c) is the effect diagram processed by the NMS algorithm.
图 8 无纺布疵点检测系统总体组成, 左边为设计示意图, 右边为现场实物图. 设计图中的(a)代表工业控制机集群, (b) 代表高速工业摄像头, 摄像头发出的的蓝色光线代表摄像头的视野. 下方的摄像头组用于预检测, 上方的摄像头组用于精确检测.
Fig. 8 The overall composition of the non-woven defect detection system, the left is the design schematic diagram, and the right is the real machine. In the design diagram (a) represents the industrial control machine cluster, (b) represents the high-speed industrial camera, and the blue light from the camera represents the camera's field of view The lower camera group is used for pre-detection, and the upper camera group is used for precise detection.
表 1 最大稳定极值区域算法选取候选点
Table 1 Maximum stable extreme value region algorithm to select candidate points
输入: 灰度图像阈值 ${{\rm{I}}_g}$ ,$\tau \in {{\mathbb{R}}^{\rm{ + }}}$ .输出: 候选点坐标 $P$ .步骤1. 求解灰度图像的最大值 ${t_{\max }} = \max ({I_g})$ 与最小值${t_{\min }} = \min ({I_g})$ , 最大稳定区域初始化为${I_s} = {\rm{O}}$ , 初始化阈值$t{\rm{ = }}{t_{\min }} + 1$ .步骤2. 取阈值 $t - 1$ ,$t$ ,$t + 1$ . 其中$t \in \{ {t_{min}},{t_{min}} + 1,$ ${t_{min}} + 2, \cdots ,{t_{max}}\} $ , 对灰度图像${I_g}$ 进行阈值操作得到二值图像${B_{t - 1}} = \left( {{I_g} < t - 1} \right)$ ,${B_t} = \left( {{I_g} < t} \right)$ ,${B_{t + 1}} = $ $({I_g} < t + 1)$ .步骤3. 对 ${B_{t - 1}}$ ,${B_t}$ ,${B_{t + 1}}$ 求联通域$\{ Q_{t - 1}^i,Q_t^i,Q_{t + 1}^i|i = 1,2, \cdots ,M\}$ .步骤4. 计算联通域 $Q_{t - 1}^i$ ,$Q_t^i$ ,$Q_{t + 1}^i$ 的面积$A_{t - 1}^i$ ,$A_t^i$ ,$A_{t + 1}^i$ .步骤5. 如果联通区域随阈值的变化率 $|A_{t + 1}^i - A_{t - 1}^i|/$ $|A_t^i| \leqslant \tau $ , 则此区域为稳定区域, 使${I_s} = {I_s} \cup Q_t^i$ .步骤6. 如果 $t < {t_{\max }}$ , 则$t$ 自增1, 重复步骤2-5.步骤7. 得到最大稳定区域 ${I_s}$ , 对其求四邻域联通域, 其联通域的质心坐标即为候选点坐标$P$ .表 2 非极大值抑制算法
Table 2 Non-maximum suppression algorithm
输入: 模型预测输出 $n$ 个外接矩形集${\mathbb{B}} = \{ {B_i}|i = 0,1,$ $\cdots ,n\}$ , 及其属于疵点的置信度集合${\mathbb{C} } = \{ {C_i}|i = 0,1, \cdots ,n\}$ ,$IoU$ 的阈值$T \in (0,1)$ .输出: 极大外接矩形集合 ${\mathbb{M}}$ .步骤1. 极大外接矩形框集合 ${\mathbb{M}} = \emptyset $ .步骤2. 求解疵点置信度集合 ${\mathbb{C}}$ 中的最大值${C_m}$ , 及其对应的外接矩形${B_m}$ .步骤3. 将 ${B_m}$ 存入极大外接矩形集合中${\mathbb{M}} \leftarrow {\mathbb{M}} \cup {B_m}$ , 并将${B_m}$ 和${C_m}$ 从集合中剔除${\mathbb{B}} \leftarrow {\mathbb{B}} - {B_m}$ ,${\mathbb{C}} \leftarrow $ ${\mathbb{C}} - {C_m}$ .步骤4. 计算矩形框集合中元 ${B_i} \in {\mathbb{M}}$ 与${B_m}$ 之间的$IoU$ , 记为$Io{U^i}$ .步骤5. 如果 $Io{U^i} > T$ , 则将${B_m}$ ,${C_m}$ 从集合中剔除${\mathbb{B}} \leftarrow {\mathbb{B}} - {B_m}$ ,${\mathbb{C}} \leftarrow {\mathbb{C}} - {C_m}$ .步骤6. 如果 ${\mathbb{B}} \ne \emptyset $ 且${\mathbb{C}} \ne \emptyset $ , 则重复步骤2-5.步骤7. 返回极大外接矩形集合 ${\mathbb{M}}$ .表 3 训练超参数配置
Table 3 Training hyperparameter configuration
参数类型 取值 输入尺寸 48×48 训练轮数 100 学习率 $1{e^{ - 4}}$ 衰减因子 0.99 批次大小 8 优化器 Adam[34] 表 4 硬件设备参数
Table 4 Hardware device information
硬件类型 参数 摄像头 像素 分辨率 物距 视角 500w 800*600 20-10 m 75度 控制器 处理器 内存 固态 显卡 i5 8G 64G Intel HD 交换机 端口数 端口参数 8个 1000M自适应RJ45端口 表 5 模型在不同阈值下的检测精度和召回率
Table 5 The detection accuracy and recall rate of the model under different thresholds
模型 指标 $\tau {\rm{ = 20}}$ $\tau {\rm{ = 30}}$ $\tau {\rm{ = 40}}$ 预检测 精度 73.0% 80.5% 86.4% 召回率 100.0% 100.0% 99.3% 精确模型 精度 86.1% 94.5% 98.4% 召回率 100.0% 100.0% 99.7% 表 6 模型预测速度测试
Table 6 Model prediction speed test
模型 48×48候选区域 2400×600无纺布图像 预检测模型 3.2 ms 46.8 ms 精确检测模型 13.9 ms 213.3 ms 表 7 不同类型疵点的检测精度和召回率测试
Table 7 Testing accuracy and recall rate of different types of defects
疵点类型 疵点尺寸 召回率 精度 点状瑕疵 0.3 mm 100% 98.8% 0.1 mm-0.3 mm 99.6% 98.3% 丝状瑕疵 0.1 mm以上 98.8% 98.6% 0.05 mm-0.1 mm 96.8% 97.2% -
[1] 周树照. 论无纺布的现状与发展. 山东工业技术, 2017, (20): 273Zhou S Z. On the status quo and development of non-woven fabrics. Shandong Industrial Technology, 2017, (20): 273 [2] De Baar H J W. von Liebig's law of the minimum and plankton ecology (1899–1991). Progress in Oceanography, 1994, 33(4): 347−386 doi: 10.1016/0079-6611(94)90022-1 [3] 陶显, 侯伟, 徐德. 基于深度学习的表面缺陷检测方法综述. 自动化学报, 2020, 1-19[2020-06-21].Tao X, Hou W, Xu D. A survey of surface defect detection methods based on deep learning. Acta Automatica Sinica, 2020, 1−19[2020-06-21]. [4] 刘伟斌, 郑力新, 周凯汀. 采用频域滤波的织物疵点检测方法. 华侨大学学报(自然科学版), 2017, (4): 562−566Liu W B, Zheng L X, Zhou K T. Detection method of fabric defects based on frequency domain filtering. Journal of Huaqiao University(Natural Science), 2017, (4): 562−566 [5] 杨亚, 薛云灿, 沙伟, 等. 基于正交小波分解的织物疵点检测. 微处理机, 2015, 36(3): 46−49 doi: 10.3969/j.issn.1002-2279.2015.03.014Yang Y, Xue Y C, Sha W, et al. Fabric defect detection based on orthonormal wavelet decomposition. Micro- processors, 2015, 36(3): 46−49 doi: 10.3969/j.issn.1002-2279.2015.03.014 [6] 曹媛媛, 杨波, 等. 基于分形纹理特征和小波变换的网状纹理检测方法. 自动化学报, 2007, 33(7): 688−692Cao Y Y, Yang B. Netlike texture detection method using fractal texture features and wavelet transform. Acta Automatica Sinica, 2007, 33(7): 688−692 [7] 吴秀永, 徐科, 徐金梧. 基于Gabor小波和核保局投影算法的表面缺陷自动识别方法. 自动化学报, 2010, 36(3): 438−441 doi: 10.3724/SP.J.1004.2010.00438Wu X Y, Xu K, Xu J W. Automatic recognition method of surface defects based on gabor wavelet and kernel locality preserving projections. Acta Automatica Sinica, 2010, 36(3): 438−441 doi: 10.3724/SP.J.1004.2010.00438 [8] 王传桐, 胡峰, 徐启永, 等. 采用Gabor滤波簇和等距映射算法的织物疵点检测方法. 纺织学报, 2017, 38(03): 162−167Wang C T, Hu F, Xu Q Y, et al. Detection of fabric defects based on Gabor filters and Isomap. Journal of Textile Research, 2017, 38(03): 162−167 [9] 孙君顶, 李欣, 盛娜, 等. 基于MBLBPV算法的布匹瑕疵检测方法. 测控技术, 2019, 38(1): 71−76Sun J D, Li X, Sheng N, et al. Fabric defect detection based on multi-scale block binary patterns variance. Measurement & Control Technology, 2019, 38(1): 71−76 [10] 张沐光, 宋执环. LPMVP算法及其在故障检测中的应用. 自动化学报, 2009, 35(6): 766−772 doi: 10.3724/SP.J.1004.2009.00766Zhang M G, Song Z H. LPMVP algorithm and its application to fault detection. Acta Automatica Sinica, 2009, 35(6): 766−772 doi: 10.3724/SP.J.1004.2009.00766 [11] 董蓉, 李勃, 徐晨. 应用积分图的织物瑕疵检测快速算法. 纺织学报, 2016, 37(11): 141−147Dong R, Li B, Xu C. Fast fabric defect detection algorithm based on integral image. Journal of Textile Research, 2016, 37(11): 141−147 [12] 杨晓波. 基于GMFR模型的统计特征畸变织物疵点识别. 纺织学报, 2013, 34(4): 137−142 doi: 10.3969/j.issn.0253-9721.2013.04.027Yang X B. Fabric defect detection of statistic aberration feature based on GMRF model. Journal of Textile Research, 2013, 34(4): 137−142 doi: 10.3969/j.issn.0253-9721.2013.04.027 [13] 李良福, 马卫飞, 李丽, 陆铖. 基于深度学习的桥梁裂缝检测算法研究. 自动化学报, 2019, 45(9): 1727−1742Li L F, Ma W F, Li L, et al. Research on detection algorithm for bridge cracks based on deep learning. Acta Automatica Sinica, 2019, 45(9): 1727−1742 [14] 金侠挺, 王耀南, 张辉, 等. 基于贝叶斯CNN和注意力网络的钢轨表面缺陷检测系统. 自动化学报, 2019, 45(12): 2312−2327Jin X T, Wang Y N, Zhang H, et al. DeepRail: Automatic visual detection system for railway surface defect using bayesian CNN and attention network. Acta Automatica Sinica, 2019, 45(12): 2312−2327 [15] 王孟涛, 李岳阳, 杜帅. 基于机器视觉的疵点检测方法的研究进展. 现代纺织技术, 2019, 27(05): 57−61Wang M T, Li Y Y, Du S. Research progress of defect detection method based on machine vision. Advanced Textile Technology, 2019, 27(05): 57−61 [16] Zhao Y, Hao K, He H, et al. A visual long-short-term memory based integrated CNN model for fabric defect image classification. Neurocomputing, 2020, 380: 259−270 doi: 10.1016/j.neucom.2019.10.067 [17] DAGM 2007 Datasets[Online], available: https://hci.iwr.uni-heidelberg.de/node/3616, February 27, 2018 [18] Kampouris C, Zafeiriou S, Ghosh A, et al. Fine-grained material classification using micro-geometry and reflectance. European Conference on Computer Vision. Springer, Cham, 2016: 778−792. [19] Kylberg G. Kylberg Texture dataset v.1.0. centre for image analysis, Swedish University of Agricultural Sciences and Uppsala University, 2011. [20] Çelik H İ, Dülger L C, Topalbekiroğlu M. Development of a machine vision system: real-time fabric defect detection and classification with neural networks. The Journal of The Textile Institute, 2014, 105(6): 575−585 doi: 10.1080/00405000.2013.827393 [21] 车翔玖, 刘华罗, 邵庆彬. 基于Fast RCNN改进的布匹瑕疵识别算法. 吉林大学学报 (工学版), 2019, 49(6): 2038−2044Che X J, Liu H L, Shao Q B. Fabric defect recognition algorithm based on improved Fast RCNN. Journal of Jilin University(Engineering and Technology Edition), 2019, 49(6): 2038−2044 [22] Liu Z, Liu X, Li C, et al. Fabric defect detection based on faster R-CNN. Ninth International Conference on Graphic and Image Processing (ICGIP 2017). International Society for Optics and Photonics, 2018, 10615: 106150A [23] 邵鑫玉, 华继钊. 基于机器视觉的无纺布缺陷自动检测系统. 计算机科学, 2014, 41(S1): 487−489Shao X Y, Hua J Z. Automatic detection system of fabric defects based on machine vision. Computer Science, 2014, 41(S1): 487−489 [24] 李明, 景军锋, 李鹏飞. 应用GAN和Faster R-CNN的色织物缺陷识别. 西安工程大学学报, 2018, 32(06): 663−669Li M, Jing J F, Li P F. Yarn-dyed fabric defect detection based on GAN and Faster R-CNN. Journal of Xi’an Polytechnic University, 2018, 32(06): 663−669 [25] Matas J, Chum O, Urban M, et al. Robust wide-baseline stereo from maximally stable extremal regions. Image and Vision Computing, 2004, 22(10): 761−767 doi: 10.1016/j.imavis.2004.02.006 [26] 贺智明, 彭亚楠. 基于深度学习的织物疵点检测研究进展. 毛纺科技, 2019, 47(08): 83−88He Z M, Peng Y N. Progress in fabric defect detection based on deep learning. Wool Textile Journal, 2019, 47(08): 83−88 [27] Huang G, Liu Z, Van Der Maaten L, et al. Densely connected convolutional networks. Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 4700−4708. [28] Deng J, Guo J, Zhou Y, et al. Retinaface: Single-stage dense face localisation in the wild. arXiv preprint arXiv: 1905.00641, 2019. [29] Lin T, Dollár P, Girshick R, et al. Feature pyramid networks for object detection. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. 2017: 2117−2125. [30] Redmon J, Farhadi A. Yolov3: An incremental improvement. arXiv preprint arXiv: 1804.02767, 2018. [31] Zhou D, Fang J, Song X, et al. ou loss for 2D/3D object detection. 2019 International Conference on 3D Vision (3DV). IEEE, 2019: 85−94. [32] Zheng Z, Wang P, Liu W, et al. Distance-IoU loss: faster and better learning for bounding box regression. arXiv preprint arXiv: 1911.08287, 2019. [33] 童艳, 孙君亮, 王鹏宇. 实时测控数据处理集群软件测试系统设计. 自动化技术与应用, 2019, 38(06): 11 doi: 10.3969/j.issn.1003-7241.2019.06.003Tong Y, Sun J L, Wang P Y. Design of test system for real-time measurement and control data processing cluster. Techniques of Automation and Applications, 2019, 38(06): 11 doi: 10.3969/j.issn.1003-7241.2019.06.003 [34] Kingma D P, Ba J. Adam: A method for stochastic optimization. arXiv preprint arXiv: 1412.6980, 2014. [35] Liu W, Anguelov D, et al. Ssd: Single shot multibox detector. European Conference on Computer Vision, 2016: 21−37. -

计量
- 文章访问数: 52
- HTML全文浏览量: 22
- 被引次数: 0