AUTOMATION TECHNOLOGY FOR FABRIC INSPECTION SYSTEM

Transcription

1 AUTOMATION TECHNOLOGY FOR FABRIC INSPECTION SYSTEM Chi-ho Chan, Hugh Liu, Thomas Kwan, Grantham Pang Dept. of Electrical and Electronic Engineering, The University of Hong Kong, Pokfulam Road, Hong Kong INTRODUCTION In order to establish good company image and be more competitive in the marketplace, inspection system can ensure the optimum quality of product on hand and help improve the product quality. Therefore, a lot of factories pay attention to improve the inspection system, especially in textile industry. Human inspection is generally used in textile industry because the cost is very low. However, the disadvantage is that the nature of work is very dull and repetitive, which leads to high inspection error. The automated inspection by machine, on the other hand, can solve this shortcoming. There are more advantages in applying automation technology for inspection system. First, 100% inspection can be obtained. Secondly, it eliminates high inspection error due to human frailty. Thirdly, it can save the labor cost, and reduce the demand for highly skilled inspectors. This paper describes the development of an Automated Visual Inspection (AVI) system for weaving defect detection based on image processing and recognition algorithms. The literatures contain many articles to illustrate a great variety of algorithms dedicated to the description or classification the weaving defect[1-3]. The purpose of the developed system is used to fully inspect the fabric and weaving products from a textile factory. The location, size and image of the defect are recorded in the system. After the inspection process, The product will be graded in terms of severity and the detailed report will be printed. In addition to the use of standard image processing functions for enhancing and modifying the digital image, the paper will describe the techniques from neural network for classifying the weaving defect. The aim is to obtain saving of manpower and time as well as increased accuracy in the inspection process. This project was proposed from a general discussion with a Managing Director of a local textile manufacturer about the needs and problems in the textile weaving industry. There are indeed such automatic fabric inspection systems in the market (e.g. I-Tex, Elbit Vision Systems Ltd., Israel; SICK Scan, Erwin Sick GmbH, Germany; SCANTEX, Sam Vollenweider Ltd., Switzerland). However, they are very expensive and most local manufacturers cannot afford to purchase them. For example, one complete set of such machinery called I-Tex would cause several million HKD. This proposal aims to develop a low-cost alternative for the local manufacturers. Also, it is high time for the Government to realise the importance of developing and applying high-tech in Hong Kong. There is enough expertise locally and we are committed to developing an automated fabric inspection system comparable to those in the market, for the benefits of the textile and garment industries in Hong Kong. Our success is based on the followings: (1) use of the latest low-cost but high-performance computer hardware and CCD digital cameras; (2) use of advanced software techniques for image processing and defect detection; (3) years of experience and research at the Industrial Automation Laboratory at The University of Hong Kong. 2. REQUIREMENTS Although significant advances have been made in the field of computer vision, image processing and pattern recognition in recent years, it is still not possible to design a flexible, cost effective, off-the shelf

2 visual inspection for all types of industrial applications. This is due to the need to consider the individual requirements of each application. Regarding an AVI system on weaving defects, the requirements are as follows: (1) Detection of defects on fabric product as it moves through the image acquisition unit. The irregularities are detected as defects. (2) The image data must be processed at a sufficiently high rate. (3) The defects are grouped according to size, direction, shape and graded in terms of severity. It should be mentioned that the defects can come in complicated and irregular shapes. (4) The image acquired may be noisy. (5) Defect locations and their images can be displayed on monitor in real time. (6) Defect information is stored and used for subsequent analysis. (7) Marking unit and alarm can be activated. (8) Inspection report is prepared and printed. Digital images are obtained as a standard camera scanned through the fabric product. Image cleaning and modification is carried out for enhancement. Segmentation is then performed and the resultant image is used for analysis. Geometric features are obtained which is then used for pattern matching and classification. In this paper, a multilayer feedforward neutral network is presented. The general hardware implementation is discussed as well. 3. THREE LEVELS OF DIGITAL IMAGE PROCESSING ON DEFECTS The low level operations involve thresholding, edge detection, line and region finding. The following figure shows the captured image of a piece of fabric with defect on it. Figure 3.1 Image of a piece of fabric After thresholding, the image would appear as the following figure. Figure 3.2 Fabric image after thresholding The regions of the defects are found which is shown in the following figure.

3 Figure 3.3 Fabric image after finding the region of defect Individual defect and the overall region covered by all the defects are windowed. The medium level operations involve feature extraction. Five features are used in our tests. The first one is the height of the overall defects window. The second feature is the width of the overall defect window. The third feature is the ratio of total defect area to the overall window area. The forth feature is the number of defect occurs in the overall defects window. The fifth feature is the ratio of the smallest defect area over the largest defect area. Finally, the high level operations consist of recognition and classification. For this part, neural network is used to conduct a test which will be described in a later section. 4. ANALYSIS OF THE DEFECTS Twenty-four fabric images with different defect were captured and processed. The following table shows the information about the fabric color, lighting, process before thresholding, threshold values, defect and noise characteristic. Fabric Color Lighting Threshold values Defect Shape & Color Noise shape & Color Defect 1 (Uneven Yarn) Defect 2 (Thick Yarn) Yellow Back Light 158 Thin, short, horizontal black lines Deep Blue Few black lines of cloth texture Direct light 114 A long, white, horizontal Many Short white line line broken into segments segments due to cloth texture Defect 3 (Slubs) Yellow Back Light 165 A thicker, longer, horizontal black lines compared to defect1 Defect 4 (Trash) Yellow Back Light 162 Long, black, horizontal lines but broken Defect 5 (Coloured Yarn) White Back Light 127 A thick, short, horizontal black line Defect 6 (Dirty White Back Light 177 A long, black, horizontal Yarn) line broken into many segments Defect 7 Yellow Back Light 158 Short, horizontal black (Hessian lines Contamination) Defect 8a (Wood Pattern) Defect 9 (Nispick) Defect 10 (Double Weft) Yellow Back Light 154 Thin, short, black lines formed into groups Yellow Direct light 174 Very short, thin black horizontal lines forming a vertical line Yellow Back Light 154 A long horizontal black line Few short black lines due to cloth texture Few short black lines due to cloth texture nil Black dots and lines due to cloth texture Black dots due to cloth texture Noise and pattern difficult to distinguish Black dots and shadow due to uneven light Many thin & short black lines

4 Defect 11 (Thin Bar) Defect 12 (Thick Bar) Defect 13 (Kink) Defect 14 (Mixed Weft) Defect 15 (Wrong Draw) Defect 16 (Float) Defect 17 (Loom Fly) Defect 18 (Broken End) Defect 19 (Bad Selvage) Defect 20 (Double, Tripewarp) Defect 21 (Oil Stain) Defect 22 (Netting Multipies) Defect 23a Defect 23b (Yarn Tails) Yellow Back Light 201 White thick line segments forming few parallel horizontal lines Yellow Back Light 155 Black, thick segments Forming a horizontal bar Few white dots Thin short black lines Yellow Back Light 144 Short black horizontal lines Black dots Yellow Back Light 157 A thick, horizontal black Short black lines line Yellow Back Light Nil A long vertical line Short black lines Yellow Back Light 152 A short, thin, horizontal black line Yellow Back Light 150 Thick black short horizontal lines Yellow Back Light 112 A thin, white, long vertical line Yellow Back Light 160 Black curves and lines forming a roots pattern Yellow Back Light 169 Black areas forming a Vertical line Yellow Room Light 177 Few black areas distributed over the cloth Yellow Back Light 210 White areas concentrated in a portion Light Blue Direct light 151 Black separated dots forming two parallel vertical lines Yellow Direct light 152 Think horizontal black line segments Short thin black lines Few black dots. White short lines Many black dots Quite big black areas Big black dots Almost none Black dots Black dots Table 4.1 Fabric image segmentation, defect and noise characteristics 5. NEURAL NETWORK APPROACH FOR DEFECT CLASSIFICATION The neural network uses the knowledge and inference procedures to solve problems which usually require human expertise for the solution. Hence, it is very suitable for classifying the weaving defect Width Height Class 5 Defect Image Feature Extractor Defect Area Number of Occurance of Defect Defect area ratio Class 6 Class 10 Input Layer Internal Layer Output Layer Figure 5.1 Neural Network architecture for defect detection

5 The architecture of a neural network for defect classification is shown in figure 5.1. A multi-layer perceptions, using the back-propagation method of Rumelhart et al. [4] is used in this paper. A neural net was trained to classify fifteen defects. Backpropagation learning rules were used to adjust the weighs and biases of networks so as to minimize the sum squared error of the neural networks. Several parameters of the network were adjusted to yield the best result. The configuration of the neural network with a good performance is as follows: two layer network, twenty neurons in the hidden layer, error goal and learning rate equals 1. It has been found that there is some improvement when the number of hidden unit is increased. However, the performance would drop if the learning rate is either too large or too small. The neural network is then tested using some test samples and an accuracy of 93% is obtained. With the use of a neural network, the time needed for defect classification is very fast and is well-suited for a real-time inspection system. 6. HARDWARE IMPLEMENTATION Complete image processing usually requires a sequence of different tasks to be performed on an image. In addition, a 72-camera array will be used to collect the weaving fabric image. Such architecture would require very long computation time to process the image in the past. Fortunately, the computer network is rapidly developing in recent years. The special-purpose hardware device and processors can be connected together to achieve the real-time performance. Downton and Crookes[5] proposed the bus-based architecture for parallel processing in This parallel architecture is very useful in AVI system for weaving defect detection and it is shown in figure Hub... Hub Hub Low level operation Illumiation light box control Motor Control High Speed bus network Medium level operation High level operation File and Control Sever High Speed bus network Figure 6.1 Bus-based architecture for automatic fabric inspection This architecture is very flexible, which can include hubs, illumination light box control, motor control and multiple processors. In multi-camera systems, the cameras are connected to the processor via hubs. Hubs perform multiplexing of camera thus allowing each connected device to be accessed individually. This idea is proposed by Lehotsky[6] in Hubs are used to transfer the image data to the processors for low level image operation. After this operation, the data will pass to the mediu m level processors to extract the features. Then the features will be used to classify object and stored it in the file sever at high level network.

6 7. CONCLUSIONS This paper has described a high speed intelligent inspection system for detecting and classifying the weaving defects. The neural network approach seems to be an effective tool for classifying the weaving defect. In addition, once the network is trained, the step of classification involves just a feedforward operation which is very fast. For 100% inspection, a camera array will be used. Using hubs is the solution which allows for a large number of cameras to be connected to a single processor. The bus-based network architecture was also presented in this paper which can achieve the real time performance in processing a large volume of image data. 8. REFERENCES 1. I. M. Dar, W. Mahmood, G. Vachtsevanos, Automated pilling detection and fuzzy classification of textile fabrics. Machine vision applications in industrial inspection V. SPIE Vol pp H. C. Abril, M. S. Millan, R. Navarro, Pilling evaluation in fabrics by digital image processing. Vision system: applications. SPIE Vol pp C. Ciamberlini, F. Francini, G. Longobardi, P. Poggi, P. Sansoni, B. Tiribilli, Weaving defect detection by fourier imaging. Vision system: applications. SPIE Vol pp D. E. Rumelhart, G. E. Hinton, R. J. Williams, Learning internal representation by error propagation. Parallel distributed processing: Explorations in the microstructures of cognition, MIT Press, Cambridge, M.A. Vol. 1. pp A.Dowton, D. Crookes, Parallel architecture for image processing D. A. Lehotsky, Developments in high speed inspection using intelligent CCD cameras. Machine vision applications, architectures, and system integration V. SPIE Vol pp