Integrated Digital System for Yarn Surface Quality Evaluation using Computer Vision and Artificial Intelligence Sheng Yan LI, Jie FENG, Bin Gang XU, and Xiao Ming TAO Institute of Textiles and Clothing, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong Abstract - The evaluation method of yarn surface quality currently in use is mainly based on manual inspection. In order to resolve the inherent limitations of the human visual inspection, an intelligent evaluation system has been developed for the objective and automatic evaluation of yarn surface quality with computer vision and artificial intelligence. In this system, all yarn surface features are fully digitalized and quantitatively processed to ensure an objective evaluation of yarn surface appearance. This digital system integrates and controls the whole progress of yarn surface analysis, including the image acquisition, digital feature extraction, characteristic parameter computation and yarn quality classification, in one computer program with an interactive and friendly user interface. Besides yarn quality classification, multiple yarn surface characteristics, such as yarn diameter irregularities, yarn fault areas, foreign matters and fuzziness, can also be quantitatively obtained and visibly displayed. Keywords: Yarn evaluation, Intelligent system, Image processing, Artificial neural network 1 Introduction The grade assessment of yarn appearance quality, or so called yarn surface grading, is one of the important testing procedures in the textile industry. According to ASTM D 2255 [1], a standard test method is to wind a yarn sample on a black board using a yarn board winder and then compare the board with a series of photographic standards representing the grades A (best), B, C and D (worst), which assesses yarn surface quality with consideration of the unevenness, fuzziness, neppiness and visible foreign matter. Traditionally, the inspection is carried out by direct observation in which a skilled specialist visually compares the wound yarn sample with the grade labeled photographic standard and then judges the quality of the yarn sample according to the standard definition, as shown in Figure 1. But the method is subjective, time-consuming, and sometimes inconsistent. With the rapid development of computer technology, the image processing and artificial intelligence technologies become more widely used nowadays in textiles. During the past decades, the investigation on digital yarn analysis by using computer vision has attracted an increasing interest of researchers and some valuable research works have been carried out [2-7]. Yarn sample Yarn board winder Human visual observation Figure 1. Traditional method for yarn surface evaluation The paper is an extension of our preliminary work [7] on digital characterization and evaluation of yarn surface appearance. In this paper, we have further developed and implemented an intelligent system for the objective and automatic evaluation of digital yarn appearance quality using computer vision and artificial intelligence. The newly developed system is able to integrate and control the whole progress of yarn surface analysis, including the image acquisition, digital feature extraction, characteristic parameter computation and yarn quality classification, with an interactive and friendly user interface. Besides yarn quality classification, multiple yarn surface characteristics, such as yarn diameter irregularities, yarn fault areas, foreign matters and fuzziness, can also be quantitatively obtained in this system for further study and evaluation. In the following section, the interface and function of the integrated evaluation system for digital yarn appearance quality will be firstly introduced in Section 2. Then, in Section 3, the methodology of the system will be illustrated in details. Finally the conclusions will be given in Section 4.
Figure 2. Digital Yarn Surface Grading System 2 Digital and Visible Yarn Surface Grading System 2.1 System Interface and Function The graphics user interface (GUI) of the visible evaluation system for digital yarn appearance quality is shown in Figure 2. In this system, users can carry out the yarn image processing and appearance classification by a series of simple operations, and then the figures and relevant features of the digital yarn image can be displayed visibly. As shown in Figure 3, the main steps of the Digital Yarn Surface Grading System are, firstly, to acquire the yarn image using a commercial scanner, next to load the image into the system, then to conduct the yarn image processing, and finally to classify the yarn appearance grade. Open scanner to obtain yarn image Yarn appearance classification Load scan image into the program Online image processing Figure 3. Main steps of the Digital Yarn Surface Grading System In yarn image processing, the important processing figures and the statistical results of yarn image listed in Table 1 can be computed and shown in the system interface. Besides, the extracted features for yarn appearance classification based on the image processing can also be shown in the input feature frame of the artificial neural network region. Table 1. Visible Results Maps and figures Gray image Binary image Hairiness image Histogram Width map Saliency map Statistical results Value of thick place Value of thin place Value of neps Mean value of diameter Standard deviation of diameter Value of hairiness In addition, the system also provides online help for users, including the user guide (Demo button), flow chat of the program (Flow Chart button) and the digital yarn database for training the artificial neural network (Database button). 2.2 Digital Yarn Database A series of weaving and knitting yarns with different appearance qualities are produced by different spinning methods using different materials and spinning parameters. All these yarn samples are physically measured by Uster Tensorapid for yarn strength, Zweigle hairiness tester for yarn hairiness and Uster III tester for yarn evenness. And the digital yarn images, which are labeled for training the artificial neural network in the yarn surface grading system, are acquired by a scanner. In order to manage the yarn information, a digital yarn database management system is established based on Access database, as shown in Figure 4, which can be started by clicking Database button in the grading system. By
retrieving yarn count, yarn appearance grade or the specified ID of yarn samples (see Figure 4 (a)), the physical properties and digital images of yarn can be displayed in the system interface (see Figure 4 (b)). (a) Index and queries interface Figure 6. Original scanned yarn images with different appearance grades 3.2 Image Processing (1) Wavelet transformation for yarn hairiness extraction (b) Detailed information interface Figure 4. Yarn image sample database management system 3 3.1 Wavelet transform provides a multi-resolution analysis and is exploited to extract texture characteristics of yarn hairiness. Figure 7 shows identification results of yarn hairiness from the two yarn images. Methodology Image Acquisition In the interface of the intelligent evaluation system, clicking the Scan Yarn Image button on the left top corner can open a scanner for digital yarn image acquisition, with the main steps shown in Figure 5. High image resolution is adopted in digital yarn image acquisition to allow the accurate and consistent evaluation results. Figure 6 shows two scanned images of different grade yarns. These two samples will be used for showing the performance of the methodology in the system when analyzing different grade yarns. Scanner Yarn sample Yarn board winder Figure 5. Yarn image acquisition Computer Figure 7. Hairiness image
(2) Fast Fourier transform for yarn diameter segmentation In digital yarn image processing, fast Fourier transform (FFT), Butterworth filters and threshold method are employed to segment yarn diameter from the whole image. Firstly, the scanned color yarn image is transferred to gray image, then changed into frequency domain (FFT) for filtering the hairiness and noise using Butterworth filter. After that an automatic threshold method - Otsu method which chooses the threshold to minimize the interclass variance of the black and white pixels, is used to get the binary image, as shown in Figure 8. 1.5 0.5 2 x 104 1 0 0 2 4 6 8 10 12 14 14000 12000 10000 8000 6000 4000 2000 0 0 5 10 15 20 25 30 35 Figure 9. Histogram of yarn diameter Figure 8. Binary image 3.3 Feature Extraction and Classification (1) Yarn diameter statistics Statistical measurement is employed for feature extraction of yarn diameter. Figures 9 to 11 show the histogram, width map and saliency map of yarn diameter. Saliency map [8] is an visual attention method which can topographically identify the visual saliency or distinguished areas of a visual scene by considering the centre-surround contrasts in terms of visual features, including intensity, color and orientations. Here, saliency map is used for yarn fault (abnormal region) detection. Figure 10. Width map of yarn diameter
shows the important characteristic results with an interactive and friendly user interface. This computerized technology is potential for commercialization and can be applied in textile testing laboratories and spinning mills for yarn surface quality control and assurance. 5 Acknowledgments The authors wish to acknowledge the funding support from the Hong Kong Polytechnic University for the work reported here. Miss Li SY and Mr Feng J would also thank the Hong Kong Polytechnic University for providing them with postgraduate scholarships. 6 References [1] Standard Test Method for Grading Yarn for Appearance ; ASTM D 2255/D2255M - 09, pp. 1-5, 2009. Figure 11. Saliency map based yarn fault detection (2) Artificial neural network for yarn appearance classification In the interface of the developed evaluation system, the Yarn Appearance Classification button is used to employ an artificial neural network to classify and grade yarn surface quality. Based on the above yarn image processing, 18 features including statistical results of yarn diameter and texture characteristics of yarn hairiness, are extracted as input parameters for the artificial neural network. With over 400 training samples in different yarn linear densities (20Ne-80Ne) and appearance grades, the result shows around 87% of over 170 testing samples can be correctly classified by using this artificial neural network [7]. 4 Conclusions A novel integrated intelligent evaluation system was developed to replace the conventional manual inspection for the objective and automatic evaluation of yarn surface appearance with computer vision and artificial intelligence. In the developed system, some recent advances in digital processing and computer science, such as saliency map analysis, wavelet transform and artificial neural network, are developed and incorporated to fully extract the yarn surface characteristic features and then to classify and grade yarn surface qualities based on the digital features. This system integrates the whole progress of yarn surface analysis and [2] D. Semnani, M. Latifi, M.A. Tehran, B. Pourdeyhimi and A.A. Merati. Grading of Yarn Surface Appearance Using Image Analysis and an Artificial Intelligence Technique ; Textile Research Journal, vol. 76, no. 3, pp. 187-196, 2006. [3] X. Zhou. Study on Yarn Blackboard by Digital Image Processing Method ; Modern Applied Science, vol. 1, no. 4, pp. 107-111, 2007. [4] J. Liu, Z. Li, Y. Lu and H. Jiang. Visualisation and Determination of the Geometrical Parameters of Slub Yarn ; FIBRES & TEXTILES in Eastern Europe, vol. 18, no. 1, pp. 31-35, 2010. [5] R. Pan, W. Gao, J. Liu and H. Wang. Recognition the Parameters of Slub-yarn Based on Image Analysis ; Journal of Engineered Fibers and Fabrics, vol. 6, no. 1, pp. 25-30, 2011. [6] H.C. Lien and S. Lee. A Method of Feature Selection for Textile Yarn Grading Using the Effective Distance between Clusters ; Textile Research Journal, vol. 72, no. 10, pp. 870-878, 2002. [7] Z. Liang, B.G. Xu, Z.R. Chi and D.G. Feng. Intelligent Characterization and Evaluation of Yarn Surface Appearance Using Saliency Map Analysis, Wavelet Transform and Fuzzy ARTMAP Neural Network ; Expert Systems with Applications, vol. 39, no. 4, pp. 4201-4212, 2012. [8] L. Itti, C. Koch and E. Niebur. A Model of Saliency- Based Visual Attention for Rapid Scene Analysis ; IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 20, no. 11, pp. 1254-1259, 1998.