A High-Precision Fusion Algorithm for Fabric Quality Inspection

Transcription

1 nd International Conference on Computer Engineering, Information Science and Internet Technology (CII 2017) ISBN: A High-Precision Fusion Algorithm for Fabric Quality Inspection YUJUN CHEN, LIMING WU, XIN LI and YANAN ZHAO ABSTRACT In order to improve the accuracy of fabric texture defect detection, a combination of algorithm with Grabcut and convolutional neural network is proposed. Firstly, a segmentation algorithm based on grab-cuts is used to locate and segment the defects in fabric images accurately. Secondly, the sample images are expanded to increasing the number of the training samples. And then, the convolutional neural network is optimized to learn the features of the fabric defects more efficiently, which make it suitable for fabric defect recognition and classification. Experimental results shows that compared with other traditional algorithms, our model gets better performance with high accuracy of fabric defect detection. KEYWORDS Fabric defect detection; Grabcut algorithm; convolutional neural network. INTRODUCTION In textile manufacturing industry, the defect detection of the fabric is an indispensable part in the quality control process and the distribution of defects directly affect the quality evaluation of the fabric. The normal fabric has regular and orderly texture structure, but the defects in fabrics destroy the original texture of the fabric and affect the quality of the product. The complex texture structure of fabrics has great background interference to improve the accuracy of fabric defect detection. Therefore, in fabric detection field, the intelligent fabric detectors based on image recognition and classification has been the research hotspot, attracting a large number of experts and scholars [1-4]. In recent years, a lot of new innovative algorithms have been put forward. Malek A S et al. [5] used fast Fourier transform and cross correlation techniques to analyze the fabric image in the spatial domain, and obtain the fabric structure characteristics to detect the defects. Alper Selver M. et al.[6] proposed a texture detection based on texture statistics and gradient search, combined with differential histogram and cooccurrence matrix for fabric texture analysis, so as to speed up the processing speed of detection. In the Model-based inspection methods, Zhou J et al. [7] proposed a fabric defect detection algorithm based on an adaptive dictionary learning, using a linear combination of the dictionary to effectively represent the structure character of the normal fabric images. Then, Qu T et al.[8] improved the algorithm, considered the Yujun Chen, @qq.com, Liming Wu, jkyjs@gdut.edu.cn, Xin Li, @qq.com, Yanan Zhao, @qq.com School of mechanical and electrical engineering, Guangzhou , China; 575

2 problem of different defect size and proposed a double-scale complete dictionary, which improved the self-adaptability of the detector. Since 2012, Alex Krizhevsky s research team won the ILSVRC-2012 competition [9], deep learning caused a great shock in the field of computer vision with its strong learning ability, and sparked an upsurge in research on Artificial Intelligence. At present, it has been widely used in face recognition [10-11], traffic sign recognition [12], target recognition [13-14], etc. Due to the complex characteristic of fabric texture, the application of deep learning in fabric defect detection has aroused extensive attention of experts and scholars. For example, Jing Jun Feng [15] presents a yarn-dyed fabric detection algorithm based on convolution neural network. It designed a deep convolution neural network and used the back-propagation algorithm to adjust the network s weight and dictionary, and finally used the Meanshift algorithm to segment the fabric defects, obtaining the good detection results. However, the deep convolution neural network has too many parameters, so it is easy to cause over-fitting problem in the training process because of the insufficiency of the fabric detection database. Aiming at these difficulties, this paper presents a fabric defect detection method which combines Gradcut algorithm [16] and optimized convolution neural network. ALGORITHM FRAMEWORK Fabric defect detection mainly includes three processes: image segmentation of the ROI, dataset expansion and defect classification: 1) Firstly, the noise should be removed from the picture in the pre-processing stage. Then the Grabcut algorithm is used to segment the target area. 2) Secondly, the sample images are expanded by being rotated, stretched and other ways for increasing the number to set up the samples database; 3) finally, input the samples to the convolution neural network for model training to achieve the classification of fabric defects. The specific realization process is shown in Figure 1. Figure 1. The flow Chart of Detection Algorithm. Figure 2. Grabcut segmentation of target area. 576

3 LOCALIZATION OF THE ROI The image segmentation algorithm based on graph theory has superior performance in texture image segmentation [17]. Aiming at the problem of locating the ROI of the fabric defects, this paper proposes a Grabcut algorithm to locate the target defect. The advantages are as follows: 1) The GMM (Gaussian Mixture Model) is used instead of histogram to estimate the color probability distribution. And iterative method is used to estimate Gauss mixture model parameters; 2) the user interaction mode is simplified by using incomplete labeling method. The energy function of the Grabcut method is defined as: E( k, z) U( k, z) V( z) (1) U( k, z logp ( zn n, kn, ) log ( n, kn) ) (2) ( m, n) C exp z z 2 V ( z) n m m n Where U ( ) is data item, ( ) n 0,1, with 0 for background and 1 for foreground; is the gray histogram of image foreground and background; z is an array of the gray value of the image, z z,..., z n,..., z ) ; V is smooth item; is opacity, ( 1 N k ( k 1,..., k n,..., k N ) is the GMM label of each pixel, p ( ) is the Gauss probability distribution, and ( ) is the mixed weight coefficient. When the energy function achieves the global minimum value argmin E( k, z), it is the segmentation result obtained by the Grabcut algorithm. Compared with the traditional image segmentation algorithm, the ROI of the fabric defect provided by the Grabcut algorithm, as shown in Fig. 2, save the texture feature completely, which lays the foundation for further classification and recognition for convolution neural network. (3) DESIGN OF THE DEEP LEARNING MODEL Convolutional Neural Network Convolutional Neural Network (CNN) is a kind of artificial neural network. Sparse connection and weight sharing makes the training parameters of the convolutional neural network decrease sharply, reducing the complexity of the network [18]. One disadvantage of the traditional classifier is that it is difficult to extract high-level features sufficiently to achieve high-precision identification. Based on the network structure of AlexNet, this paper designs a CNN network structure suitable for fabric defect detection. As shown in Fig.3, compared with the AlexNet network, the network simplifies the number of convolution layers and the number of neurons in the full connection layer, greatly reducing the parameters in the network. 577

4 Figure 3. Structure of the Network. Hole Stain Broken warp broken weft warp hanged knob Figure 4. Samples of fabric defects. The model consists of three convolutions and three pooling layers. The size of the convolution kernel is 9 9, 5 5, 3 3, and the size of the pooling factor is 4 4, 3 3, 2 2, respectively. Data Expansion Training network parameters need a large number of sample data, so it is easy to cause over-fitting when the number of samples is small. And there are not enough quantity of fabric defect samples in reality to choose. Therefore, this paper has carried out the necessary manual data expansion before the acquired sample images input network training. The array representation of the training data is changed and the label remains unchanged, resulting in a certain number of sub-images, by using methods such as rotation, random stretch, and so on. For example, the rotation angle is set to - 10 and 10 degrees, the stretching factor is set to 0.8 and 1.2. EXPERIMENTS AND RESULTS ANALYSIS Experimental Platform and Data Preparation The original image data set is selected from the TILDA database, totaling 1200 pictures.figure 4 illustrates several common features of fabric defects. Selecting 70% samples of the database randomly to build traning set, and 30% as verification set. The image pixels used in the experiment are Comparison of Different Methods of Experiment Four kinds of fabric defect detection methods are selected to compare with the algorithm we proposed. There are Gray Level Co-occurrence Matrix (GLCM), Gaborbased filter, SVM and deep convolution neural network (DCNN). The experimental results are shown in Table

5 TABLE 1. DETECTION ACCURACY OF FABRIC DEFECTS BY DIFFERENT METHODS. Detection Method Positive defect rate for different defects (%) method Hole knob Stain hanging Broken Broken warp warp weft GLCM Gabor SVM DCNN Method of the paper As seen in the results, the deep learning achieves the ideal effect in the six kinds of fabric defect detection by its strong feature learning ability. It has shown obvious advantages compared with the GLCM method, Gabor-based filter method and SVM Methods and other traditional algorithms. In particular, defects such as broken warp and broken weft which is difficult to distinguish from background texture are still successfully recognized with high accuracy in the experiment. Our algorithm makes the average accuracy of defect detection reach more than 98%, which has better performance than DCNN. In the detection time, the average detection time of the algorithm we proposed is 0.38s, which is more efficient than DCNN, and is more than 50% faster than the other three traditional algorithms, as shown in Figure 5. Figure 5. Average inspection time of different kinds of defects by different methods. Figure 6. Influence of iteration times on the accuracy of different networks. 579

6 TABLE 2. COMPARISON OF TRAINING TIME. Network structure Accuracy (%) Convergence time(s) LeNet VGG Our Algorithm The Influence of the Number of Iterations In order to further evaluate the network performance, we also design a set of experiments to test the influence of the number of iterations on the accuracy and convergence rate of the optimized convolutional neural networks. We use LeNet[19], VGG16[20] and the network structures we proposed to compare the fabric defects experiment. The test results are shown in Fig 6. From the curve of ACC in Figure 6, it is not difficult to see that LeNet is better than the proposed algorithm when the number of iterations is small, but when the number of iterations is more than 500, the classification accuracy of the proposed algorithm is higher than that of LeNet and VGG16. When the number of iterations reaches 1000, the three networks converge gradually, and the accuracy rate is 75.6%, 82.5% and 98.2% respectively. For the convergence speed, when the number of iterations reaches about 1000, the convergence time of the algorithm we proposed is similar to that of LeNet, but is more than 30% faster than VGG16, as shown in Table 2. And the also reflects the obvious advantages in ACC. Therefore, this algorithm is superior to traditional convolution neural network model in network training. SUMMARY This study provides a new idea for fabric defect detection with complex texture features. At present, the accuracy of the fabric defect detection is difficult to achieve new breakthrough. So, a new defect detection model based on Grabcut algorithm and optimized convolutional neural network is designed. The experimental result shows that the network has faster recognition speed and higher recognition accuracy. The recognition effect is better than the existing traditional machine learning method, and can detect a variety of different texture feature of fabric defects. In the next research work, we should increase the data set and further optimize the convolution neural network framework, and explore more efficient network training strategy. ACKNOWLEDGEMENTS This research was supported by:guangdong provincial science and technology plan project (2015A and 2016A ); Guangzhou science and technology plan project ( ). REFERENCES [1] Li Wenyu, Cheng Longdi. New progress of fabric defect detection based on computer vision and image processing[j]. Journal of textile research, 2014, 35(3):

7 [2] Mak K L, Peng P, Yiu K F C. Fabric defect detection using morphological filters[j]. Image and Vision Computing, 2009, 27 (10): [3] Jing Jun-Feng, Zhang Huan-huan, Li Peng-fei. Fabric Image Defect Detection Based on Method Library [J]. Journal of Donghua University (Natural Science Edition) 2013, 39(5): [4] Zhu Dandan, Pan Ruru, Gao Weidong. Fabric defect detection using characteristic spectrum of Fourier transform and correlation coefficient. Computer Engineering and Applications [J]. 2014, 50(19): [5] Malek A S, Drean J Y, Bigue L, et al. Optimization of automated online fabric inspection by fast Fourier transform (FFT) and cross-correlation [J]. Textile Research Journal, 2013, 83(3): [6] Alper Selver M, Avşar V, Özdemir H. Textural fabric defect detection using statisti-cal texture transformations and gradient search [J]. The Journal of the Textile Institute, 2014, 105(9): [7] Zhou J, Wang J. Fabric defect detection using adaptive dictionaries [J]. Textile Research Journal, 2013, 83(17): [8] Qu T, Zou L, Zhang Q, et al. Defect detection on the fabric with complex texture via dual-scale over-complete dictionary [J]. The Journal of the Textile Institute, (6): [9] Krizhevsky A, Sutskever I, Hinton G E. ImageNet classiflcation with deep convolutional neural networks. In: Advances in Neural Information Processing Systems 25. Lake Tahoe, Nevada, USA: Curran Associates, Inc., [10] Khalajzadeh H, Mansouri M, Teshnehlab M. Face recognition using convolutional neural network and simple logistic classifier. In: Proceedings of the 17th On-line World Conference on Soft Computing in Industrial Applications. Switzerland: Springer International Publishing, [11] Li H, Lin Z, Shen X, et al. A convolutional neural network cascade for face detection[c]// Computer Vision and Pattern Recognition. IEEE, 2015: [12] Qian R, Zhang B, Yue Y, et al. Robust Chinese traffic sign detection and recognition with deep convolutional neural network [C] // International Conference on Natural Computation. IEEE, 2016: [13] He K M, Zhang X Y, Ren S Q, Sun J. Deep residual learning for image recognition. arxiv: , [14] Bell S, Zitnick C L, Bala K,et al. Inside-outside net: Detecting objects in context with skip pooling and recurrent neural networks. In CVPR 2016 [15] Jing Junfeng, Fan Xiaoting, Li Pengfoi, Hong Liang. Yarn-dyed fabric defect detection based on deep-convolutional neural network [J]. Journal of Textile Research. 2017, 38(2): [16] Chen D, Chen B, Mamic G, et al. Improved Grab Cut segmentation via GMM optimization[c]. Proc of the 2008 Int Conf on Digital Image Computing: Techniques and Applications. Washington DC: IEEE Computer Society, 2008: [17] Liu Song-Tao1, Yin Fu-Liang. The Basic Principle and Its New Advances of Image Segmentation Methods Based on Graph Cuts [J]. Acta Automatica Sinica, 2012, 38(6):

8 [18] Fukushima K. Neocognitron: a self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position. Biological Cybernetics, 1980, 36(4): [19] Y. LeCun, L. Bottou, Y. Bengio, et al. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998, 86(11):2278~2324. [20] Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image Recognition [J]. Computer Science,