英文原文.pdf

Research Article Surface defect detection in tiling Industries using digital image processing methods: Analysis and evaluation Mohammad H. Karimi, Davud Asemani n Laboratory of Signals and Electronic Systems, Electrical and Computer Engineering Faculty, K.N. Toosi University of Technology, Shariati Avenue, Tehran 1355-16315, Iran a r t i c l e i n f o Article history: Received 17 June 2013 Received in revised form 21 October 2013 Accepted 21 November 2013 Available online 4 February 2014 This paper was recommended for publication by Mohammad Haeri Keywords: Surface defect Tiling Pattern recognition a b s t r a c t Ceramic and tile industries should indispensably include a grading stage to quantify the quality of products. Actually, human control systems are often used for grading purposes. An automatic grading system is essential to enhance the quality control and marketing of the products. Since there generally exist six different types of defects originating from various stages of tile manufacturing lines with distinct textures and morphologies, many image processing techniques have been proposed for defect detection. In this paper, a survey has been made on the pattern recognition and image processing algorithms which have been used to detect surface defects. Each method appears to be limited for detecting some subgroup of defects. The detection techniques may be divided into three main groups: statistical pattern recognition, feature vector extraction and texture/image classifi cation. The methods such as wavelet transform, fi ltering, morphology and contourlet transform are more effective for pre-processing tasks. Others including statistical methods, neural networks and model-based algorithms can be applied to extract the surface defects. Although, statistical methods are often appropriate for identifi cation of large defects such as Spots, but techniques such as wavelet processing provide an acceptable response for detection of small defects such as Pinhole. A thorough survey is made in this paper on the existing algorithms in each subgroup. Also, the evaluation parameters are discussed including supervised and unsupervised parameters. Using various performance parameters, different defect detection algorithms are compared and evaluated. then, size difference compared to ideal size is measured by a stacker; fi nally, surface defects are identifi ed by human vision and registered on the product surface with fl uorescent markers. This traditional and non-automatic grading process suffers from problems such as poor performance, non-repeatable procedure, high cost, and low speed. Industrial and unhealthy environment of product line for huma- nitarian personnel is another negative factor of manual grading. The automatic grading system would result in better perfor- mance, lower cost, and uniformity in each category of products. The current increasing demand of tile and ceramic validates the market need of automatic grading for higher production speeds 2. In modern production lines, tiles are actually classifi ed into fi ve grades based on the three above mentioned evaluation criteria, in which level fi ve is considered as losses 3. Up to now, various processing algorithms have been proposed for intelligent grading. These methods can be divided into four main categories according to the defect detection mechanism: fi ltering methods, structural techniques, statistical methods, and model-based techniques (Table 1). Filtering methods usually use mathematical translation and fi lters or pattern recognition meth- ods for defect detection. The structural approaches consist of conventional morphological image processing and edge detection algorithms. Model-based approaches include common image Contents lists available at ScienceDirect journal homepage: /locate/isatrans ISA Transactions 0019-0578/$-see front matter fax: +98 21 8846 2066. E-mail address: Asemanieetd.kntu.ac.ir (D. Asemani). ISA Transactions 53 (2014) 834844 processing models like the Auto-Regressive (AR) and Hidden Markov Models (HMM). In the statistical approaches, luminance histogram is generally used for defect detection. Statistical meth- ods are characterized by simplicity as well as low complexity 4. Because of various chemical and mechanical processes in the ceramic tile production line, diverse types of surface defects appear on the fi nal product. The defects generally have different visual patterns which are sometimes contradictory. Therefore, the desired grading system should include a variety of image proces- sing algorithms to cover different types of surface faults or defects. In this paper, the proposed algorithms for grading system in ceramic and tile production line are discussed and evaluated in terms of output quality and computational complexity. In Section 2, different types of surface defects appearing in the fabrication lines of ceramic and tiles are studied. In Section 3, different defect detection algorithms are discussed. Then, Section 4 deals with the evaluation parameters. Firstly, available measures described for evaluating defect detection algorithms are presented. Using qual- ity parameters, proposed techniques are compared. Finally, the discussions are concluded in Section 5. 2. Surface defects of ceramic and tiling Ceramic and tile products pass various chemical and mechanical stages through the production line. Production of ceramic tiles comprises eight main stages: forming, drying, glazing, baking, grading, and sorting 92 as shown in Fig. 1. Glazing defects occur in glazing and printing stages. Defects that are associated with breaking and cracks happen in the forming and baking stages. In contrast, edge defects are caused more by the transmission process from glazing lines to kiln. Also, the Pinhole defect occurs typically in kiln 3. Accordingly, surface defects can be divided into six categories with the following characteristics (Fig. 2) 5. ?Pinhole Pinhole is a quality fault appearing as small holes on the product surface. Pinhole sizes are typically less than one millimeter. Also, the holes appear with a lumber and depres- sion. This fault typically occurs during baking. ?Eclipse glaze This problem originates from accumulation of a part of glaze over a corner or part of the tile. Accumulation of glaze is usually on a few millimeters with signifi cant expansion in the region of defect. This defect appears in the glazing stage by creeping and ringing of the glaze 5. ?Crack The most common defect is the crack which occurs because of fast baking procedure with rapid increase or decrease in temperature. Cracks at the edges of the tile are mostly caused due to increasing temperature. Cracks due to decreasing temperature are also called air cracks or cold cracks and often occur because of fast baking procedure in the kiln 6. ?Blob Some patches like spot drops of water may exist on the tile surface, and are called blob defects. It occurs if humidity is not adjusted or a low sleep time is included before entering into the kiln. ?Scratch This failure occurs because of dragged color printing in some directions. This defect is often created during the transmission of products from glazing line to the kiln. ?Edge Edge defects occur most commonly in the kiln but they may be generated from other manufacturing stages 79. 3. Algorithms of defect detection for ceramic and tile products For the detection of surface defects, it is required to analyze the whole product surface. So, an image with high resolution should be fi rstly captured. The system must have appropriate lighting to obtain a suitable surface picture. According to Table 1, the defect detection algorithms may be classifi ed into four principal groups. Here, the main algorithms of each group are discussed. 3.1. Filtering approaches In the fi ltering approaches, mathematical transformations and fi lters are generally used. In this regard, both linear and nonlinear transforms may be used. The most important algorithms include the Wavelet and Counterlet transform, Independent Component Analysis (ICA) analysis, Gabor fi ltering and artifi cial neural net- works which are discussed below. 3.1.1. Wavelet transform According to the nature of multi-resolution analysis, wavelet transform has been extended for many processing applications and is sometimes known as the most powerful tool 10,11. In wavelet transform, two low-pass h and high-pass g fi lters called father and mother functions, respectively, are used in a fi lter bank way (Fig. 3) 12. In Fig. 3, the input is an n?m image and there are also four outputs of LL, LH, HL and HH with size (n/2)?(m/2). At each stage, the input image is divided into four sub-images. Wavelet transform has been used for pre-processing and texture feature extraction 13. In 2001, Kumar and Pang proposed a method of defect detection based on wavelet packet. There, the wavelet packet coeffi cients from a set of dominant frequency channels containing signifi cant information are used for the characterization of textured images. This method is useful in very soft texture changes 14. In 2005, Yang et al. applied a similar method to inspect the fabrics in textile factories for defect classifi cation using discriminative wavelet frames. For a better description of the latent structure of the textile image, adaptive wavelet frames for textile would be preferred rather than standard ones. The challenge in this method is how to select the wavelet. Also, the training stage is so dependent on the number of data points 15. Table 1 Different approaches of defect detection. ApproachProcessing algorithmReferences Filtering methodsWavelet transform1218 Countorlet transform1921 Genetic algorithm2226 ICA algorithm2730 Neural networks3235 Gabor fi lter3638 Structural algorithmsMorphology3947 Edge detection7,11,4851 Model based techniquesHidden Markov model5961 Autoregressive model6265 Statistical methodsHistogram curve66,67 Co-occurrence matrix6876 Weibull distribution7782 Autocorrelation83,84 M.H. Karimi, D. Asemani / ISA Transactions 53 (2014) 834844835 3.1.2. Contourlet transform Contourlet transformation based originally on the wavelet transform aims to overcome the weakness of selected wavelet type 16. Contourlet transformation exploits multi-resolution and space-frequency curve like the Wavelet. Contourlet transform combines pyramid laplacian with a direct two-dimensional fi lter- bank (Fig. 4). The band-pass image is converted to eight sub-images Fig. 1. General stages of fabrication in ceramic and tile factory 93. Fig. 2. Types of surface defects on the ceramic and tile products. Fig. 3. General realization of wavelet transform in image analysis. Fig. 4. General block diagram of contourlet transform. M.H. Karimi, D. Asemani / ISA Transactions 53 (2014) 834844836 from pyramid laplacian. This transform has good performance for denoising and enhancing the picture 17. In 2012, Ai et al. introduced a new method of feature extraction based on the contourlet transform and kernel locality preserving projections to extract suffi cient and effective features from metal surface images. In this study, the image information at certain direction is important for the recognition of defects, and the contourlet transform is introduced for its fl exible direction setting. The disadvantage of this method is a need for extra information of contourlet transform. However, the total classifi cation rates of surface defects of continuous casting slabs and aluminum strips are up to 93.55% and 92.5%, respectively in this work 18. 3.1.3. Genetic algorithm Genetic algorithm can fi nd a suboptimum solution for optimi- zation and searching problems 19,20. In the context of defect detection, a statistical relationship is fi rstly considered to deter- mine the pixels corresponding to surface defects. Then, the related parameters are considered as genes and genetic algorithm opti- mizes these parameters. Those parameters may represent thresh- olding point or morphological method parameters 21,22. In 2002, Zheng et al. introduced a method based on the genetic algorithm to detect surface detects. In this algorithm, morpholo- gical parameters have been used including base element and thresholding points. Though, this method is very simple, but the training stage remains a controversial challenge 23. 3.1.4. ICA ICA algorithm is a basic method of source separation 24,25. In image processing applications, the ICA algorithm generally supposes that the input image is combined of two or more independent images. ICA algorithm tries to fi nd the elementary images. In defect detection, defects are supposed to be on a foreground mixed with background pattern. Then, ICA is used to separate the foreground from the background 2628. In 2006, Tsai et al. proposed a defect detection method based on ICA. They used a constrained ICA model for designing an optimal fi lter to detect surface defects from noiseless background. The proposed algorithm requires to have defectless pattern. Also, the performance degrades in the presence of noise 27. 3.1.5. Artifi cial Neural Networks Artifi cial Neural Networks (ANN) are mostly used in machine learning and artifi cial intelligence 29,30. In image processing and defect detection systems, neural networks are used as the classi- fi er. Therefore, it is necessary to extract feature vector before applying any image to the neural network. In defect detection, feature vectors would be classifi ed into two classes of defectless and defective patterns by ANN 31. In 2008, Suyi et al. proposed an ANN for defect detection in textiles 32. Neural networks have some defects, namely, the related training process takes long time, easily trapped in local minima, which infl uence the accuracy of the algorithm. Whereas particle swarm optimization has good search ability, but in this work the Particle Swarm Optimiza- tionBack-Propagation (PSO-BP) algorithm is used for the neural network which has a fast training stage rather than the BP algorithm. 3.1.6. Gabor fi lter Gabor fi lters are the ones which have the same representation in the spatial and frequency domains. These fi lters can be obtained from combining an exponential and a Gaussian function as follows 33: Gx;y e ? x ? x02 2 y? y0 2 2 hi ? e?2ju0x?x0v0y?y0?1 where x0, y0 is the center of the receptive fi eld in the spatial domain and u0, v0 stands for the center of the fi lter in frequency domain.s and represent the standard deviations of the elliptical Gaussian along x and y. Although Gabor fi lters are not orthogonal, they cover complete information of the image and are able to choose a specifi ed frequency and direction. In surface defect detection applications, defects may be dis- criminated by applying Gabor fi lters, after a simple thresholding can result in the defect regions 34. Therefore, the Gabor fi lter is used mostly as a pre-processing stage. In 2010, Che et al. intro- duced a method for fabric defect detection in textile industry based on change in the scale of the Gabor fi lter. Their proposed method suffers from high computational complexity. Besides, this method requires a reference defectless image 35. 3.2. Structural approaches In structural methods, primary and hierarchical forms are used for defect detection leading to an intuitive procedure and com- prehensible computations. These approaches use morphological operators as well as edge detection methods for defect detection. 3.2.1. Morphological methods Morphological operators are methods for both binary and grayscale image processing based on morphology. The output value of each image pixel is determined with respect to the input pixel value and its neighbors 3638. In all morphological processings, structural elements are used. So, the appropriate choice of this element is often the most important part of the process 39. Closing operator causes some regions of the image to smoothen, which usually mixes the thin fracture and removes the small holes, and fi lls up the track in the environment 40. Morphologicaloperatorsareusuallyusedforimproving, smoothing, and noise reducing in defective images. Also, by using morphological operators, edge detection in defective images is realized 4143. In 2009, Yiu et al. proposed a new method based on morpho- logical operators to detect defects in fabric texture. In this method, Gabor wavelet was fi rstly used to design the base element of morphological methods (learning phase). The algorithm works so that the input image successively passes through opening, closing, median fi lter, closing, and thresholding stages and the defect in the output image is indentifi ed. This method is of supervised type and accounts for specifi c defects 44. 3.2.2. Edge detection algorithms Edge is defi ned as a boundary between two dissimilar regions of an image. There are many different methods for edge detection 5. One of the simplest methods for the edge detection is the gradient of image 45,46. Thresholding is the last stage in edge detection. The edge of an image can be achieved with high accuracy by choosing appropriate threshold 47. Edge detection methods are used for the detection of edge defects in the surface defect detection and image segmentation 48. In 2011, Salimian and Pourghassem proposed a method for detection of edge defect in ceramic and tile. In this method, edges are fi rstly detected using the canny algorithm. Then, the angle of corners is determined by using inner product. Finally, the corner is considered as a defective edge if the angle is about 89921 9. In 2006, Mukherjee et al. introduced an edge density-based technique for defect detection and an object-based coding approach has been applied for the storage of defective ingots. For extracting the edge density, they used the Discrete C