Automatic inspection and classification for thin-film transistor liquid crystal display surface defects based on particle swarm optimization and one-class support vector machine

Abstract

Thin-film transistor liquid crystal display surface micro-defects are difficult to be detected using traditional threshold or edge detection methods. This article puts forward a non-destructive detection method using particle swarm optimization with one-class support vector machine to inspect thin-film transistor liquid crystal display surface micro-defects. An image acquisition system is constructed to acquire the surface micro-defects images of thin-film transistor liquid crystal display. Background textures are removed by the image preprocessing algorithm based on one-dimensional discrete Fourier transform. Moreover, the wavelet transform algorithm is used to eliminate the influence of uneven illumination. Effective characteristic parameters describing thin-film transistor liquid crystal display surface micro-defects are selected by the principal component analysis method. Classification model is developed based on one-class support vector machine using radial basis function. To validate the method above, other parameter optimization algorithms, including normal algorithm, genetic algorithm, and grid search algorithm, are used to optimize the support vector machine model parameters: penalty parameter C and kernel parameter g. In contrast, particle swarm optimization is proved to get the optimal model parameter, and the recognition accuracy of 91.7% is obtained from the particle swarm optimization–one-class support vector machine model. The results indicate the proposed system and method can accurately inspect thin-film transistor liquid crystal display surface detects.

Keywords

Thin-film transistor liquid crystal display defect detection image recognition particle swarm optimization one-class support vector machine

Introduction

Thin-film transistor liquid crystal displays (TFT-LCDs) are widely used in televisions, notebook computers, mobile phones, and other electronic products because of the advantages of power saving, thin and low radiation, space saving, and so on. However, TFT-LCD surface defects have plagued the image quality of TFT-LCD, and surface defect detection of TFT-LCD has become one of the key technologies of TFT-LCD production.

There are three basic methods for inspecting TFT-LCD surface defects: artificial visual inspection method, electrical parameter detection method, and automatic visual inspection method.¹ Affected by the reasons of subjectivity, slow speed, poor stability, and others, manual visual inspection is difficult to meet the needs of large-scale production. Electrical parameter detection method is limited to detect electrical defects. Automated optical inspection (AOI) method is to obtain the TFT-LCD surface image by machine vision system and inspect defects by image processing method. AOI is the inevitable trend of the TFT-LCD surface defects detection and classification because of its high accuracy, high speed, and non-contact detection, which has become the focus of many research scholars.^2,3

According to the size of TFT-LCD surface defects, they can be divided into macro defects and micro-defects.⁴ Macro defects are relatively large and can be identified by naked eyes. Mura is the common macro defect, and there are many experimental results on it.⁵ However, micro-defects are relatively small and cannot be identified by naked eyes or common electrical parameter method. As been well known, micro-defects can be divided into pinholes defects, scratch defects, and particle defects.

The AOI process of TFT-LCD surface defects detection mainly involves three aspects: image preprocessing, image segmentation, and defects classification. Image preprocessing and segmentation belong to the category of digital image processing, and they are traditional methods and relatively mature. Defects classification belongs to the category of pattern recognition, and the statistical learning theories are introduced into the automatic classification of surface defects, which play a very important role in TFT-LCD automation production.

The methods of TFT-LCD surface defects recognition mainly include neural network classifiers, support vector machines (SVMs), dynamic Bayesian networks, and so on. These algorithms have been widely used in TFT-LCD surface defects. Neural network method improves the recognition rate by providing adequate training samples in training process, but it has the disadvantages of weak generalization ability and longer training time and can easily lead to overfitting because of a small or unrepresentative training set. SVM is derived from the theory of statistical learning theory (SLT) by Vapnik,⁶ which is designed for pattern classification and regression estimation.^7–9 SVMs have many advantages in generalization ability, global optimization, and so on compared with the traditional pattern classification and machine learning methods such as artificial neural networks and get more and more widely used. SVM has become an effective tool for TFT-LCD surface defect detection and classification, and domestic and foreign scholars have done many researches on it.^10–13

SVM is mainly to solve the problem of binary classification. In order to achieve the classification goal, the training sets need to include positive and negative sample points. However, the number of training samples of different defects is often very unbalanced in the actual defect detection process. If we still use the classification method, the classification accuracy is low. For solving this problem, the defect detection method based on one-class support vector machine (OCSVM) is proposed, which only determines whether the sample has a defect and providing a true or false result.

In our work, a non-destructive detection method using particle swarm optimization (PSO) with OCSVM is proposed to inspect TFT-LCD surface micro-defects. The feasibility of classification is considered in this proposed system.

System construction and testing process

TFT-LCD surface defect detection system mainly consists of the following components: linear Charge Coupled Device (CCD) camera, optic lens, light source, image acquisition unit, image processing software, and related components. The schematic diagram is shown in Figure 1. DALSA’s linear CCD camera (Piranha ES80-04k-40 with 8192 pixels and 7.4 µm × 7.4 µm pixel size) was used in the TFT-LCD surface defect detection system. The diameter of TFT-LCD surface defect is less than 1 mm, and the required measurement resolution is about 15 µm. The aperture lens of SIGMA’s with 105 mm focal length and F2.8 aperture can meet the resolution requirements of defect detection. Halogen linear light source is used. Camera support bracket and light source support bracket can be adjusted, and the stage moves at constant speed on the rails drove by motor. The images of TFT-LCD surface defects are collected by CCD camera, transferred to, and stored in a computer by image acquisition card.

Figure 1.

Structure block diagram of TFT-LCD surface detection system.

The detection process of TFT-LCD surface defects is shown in Figure 2. First, TFT-LCD surface image is collected by the CCD industrial camera; then, the image noise is filtered out by image preprocessing and the defective area is divided; finally, the defect feature information is extracted for the classification of TFT-LCD surface defects; and the recognition results are given.

Figure 2.

Flowchart of TFT-LCD defect detection.

Image processing

Image preprocessing

Horizontal gate line and vertical data line are arranged regularly in TFT-LCD. Thin-film transistor (TFT) is located at the intersection of gate line and data line. Periodic background texture seen in the TFT-LCD image is the part of TFT. The scanned images of three kinds of TFT-LCD surface micro-defects (pinholes, scratches, and particles) are shown in Figure 3. They all have periodic textured background.

Figure 3.

TFT-LCD surface micro-defects: (a) pinholes, (b) scratches, and (c) particles.

Figure 4(a) is the gray-scale image of defect-free line scan in Figure 3(a). As shown, the peak position is located at the liquid crystal array. In the scan imaging process, the uneven illumination of light source will result in the uneven distribution of gray peak height, but this will not cause peaks mutation. Figure 4(b) is the gray-scale image of pinhole defect line scan in Figure 3(a). The pinhole defect leads to the mutation of image gray peak, which can determine the location of a TFT-LCD surface defect.

Figure 4.

The line scan gray-scale image and its reconstruction of liquid crystal array: (a) the gray-scale image of defect-free line scan, (b) the gray-scale image of pinhole defect line scan, (c) the reconstructed image of (a), and (d) the reconstructed image of (b).

The defect image information processing in periodic texture surface is more complex than in uniform surface, because the texture background needs to be eliminated and retained the defect information. Frequency filtering method is a common method of periodic texture filtering. Images are filtered by frequency domain transform, such as Fourier transform (FT),¹⁴ discrete cosine transform (DCT),^15,16 wavelet transform (WT),¹⁷ Gabor wavelet transform (GWT),¹⁸ and so on. First, the frequency components of repetitive texture background are eliminated in frequency domain analysis; then, the image is reconstructed by corresponding inverse transform; finally, the image without repeating textured background is obtained.

In this article, the images of TFT-LCD surface defects are obtained through linear CCD camera by line scanning. To ensure the real-time system requirements, one-dimensional (1D) discrete Fourier transform (DFT) is adopted to remove the textured background; then, the image without repeating textured background is obtained the inverse discrete Fourier transform (IDFT); finally, the defects are detected by defect recognition and classification algorithm. The abnormal fluctuations of high-frequency component of 1D signal processed by IDFT are eliminated by head and tail splicing method. Figure 4(c) and (d) is the gray distribution image reconstructed through 1D DFT and head and tail splicing. Because the gray distributions are uneven in Figure 4(c) and (d), the defects cannot be directly detected by thresholding method; otherwise, the normal is mistaken for defect. This article proposes a WT method to detect defects.

In the below equations, ${\hat{f}}_{y} (x)$ is the 1D gray-scale image reconstructed by filtering out the background, and ${\hat{f}}_{y} (x)$ is decomposed into a smooth dimensional image $s_{y} (x)$ and a detail dimensional image $d_{y} (x)$ by Haar wavelet function.

s_{y} (x) = \frac{1}{2} \sum_{i} L (i) \cdot {\hat{f}}_{y} (x + i)

(1)

d_{y} (x) = \frac{1}{2} \sum_{i} H (i) \cdot {\hat{f}}_{y} (x + i)

(2)

Here, $L (i)$ and $H (i)$ denote the high-pass and low-pass analysis filters of Haar wavelet. $s_{y} (x)$ represents the smooth dimensional image, which records the approximate coefficient of original gray-scale image. $d_{y} (x)$ represents the detail dimensional image, which inherits the details of original gray-scale image. i represents the exploded scale. Here, only one layer is decomposed, that is, i = 1.

Figure 4(c) and (d) is transformed into Figure 5 by 1D Haar WT. Figure 5(a) and (c) is the smooth dimensional image $s_{y} (x)$ ; Figure 5(b) is the details of the 1D defect-free gray image, which is relatively flat. Figure 5(d) shows the details of hole defect image, which fluctuates significantly. Therefore, the method can be used for defect extraction.

Figure 5.

The results of one-dimensional Haar wavelet transform (WT): (a) s_y(x) of Figure 4(c), (b) d_y(x) of Figure 4(c), (c) s_y(x) of Figure 4(d), and (d) d_y(x) of Figure 4(d).

Defect feature extraction

The defect features of TFT-LCD surface micro-defects are extracted after segmentation and edge detection of the TFT-LCD surface defect image. The image features mainly include five feature sets: geometry feature, shape feature, first-order statistical moments feature, second-order statistical moments feature, and moment invariants. Geometry feature set contains central location, direction, perimeter and area, and so on. Shape feature mainly describes the outer boundary and the overall shape of the region. First-order statistical moments describe the gray distribution of defects. Second-order statistical moments describe the texture feature of defects. Moment invariants have the invariance characteristics of rotation, translation, and scaling.

The shape feature has a great relationship with the size or location of defects, so it cannot be directly used as the feature of defect classification. The best features of defects classification should be independent of position, shape, and deflection angle. In order to ensure the validity of feature information, reduce the calculation amount, and improve the processing speed, the principal component analysis (PCA) is used to extract features.

There are mainly three types of TFT-LCD surface micro-defects: pinholes, scratches, and particles. The characteristics of defect types are described by 20 feature values extracted, such as area, perimeter, roundness, flatness, duty cycle, the minimum bounding rectangle (MBR), aspect ratio, gray-scale range, mean gray, gray variance, gray energy entropy, first-order moments feature, second-order moments feature, and Hu moment invariants, which constitute the feature space of TFT-LCD surface micro-defects.

The complexity of recognition grows rapidly with the increase in dimensions because the characteristic values are relevant. PCA method is used to extract the main component characteristic parameters from the 20 dimensionality features in order to reduce dimensionality and redundant data. According to the standard of cumulative contribution ratio of the principal component reaching 90%, five principal component characteristic parameters are retained, which constitute the feature subspace of TFT-LCD surface micro-defects. In order to facilitate subsequent processing, the characteristic parameters must be optimized and normalized.

Defect classification algorithm

SVM principle and pattern recognition

TFT-LCD panel defect classification is actually a pattern recognition category. SVM^19–21 is a machine learning method proposed by Vapnik based on the SLT and structural risk minimization, which is different from the traditional method based on the empirical risk minimum (ERM) principle. SVM with high generalization capability has more advantage in solving the problem of small sample, nonlinear, and high-dimensional recognition.

The basic idea of SVM is to map the input samples from low-dimensional feature space to the high-dimensional feature space by kernel functions and establish the optimal classification surface in the high-dimensional feature space. Assuming that the sample collection is ${(x_{i}, y_{i}) | i = 1, 2, \dots, N}$ , $x_{i}$ is the input of sample i, $y_{i}$ is the output category of sample i, and the category label is ${1, - 1}$ , the classification discriminant function is $g (x) = w_{0}^{T} x + b_{0}$ .

Through the normalized processing, the samples closest to classification interface meet $| g (x) | = 1$ . The category interval is $M \arg in = 2 / | | w | |$ and the category output meets $y_{i} (w^{T} x_{i} + b) \geq 1$ . For the case of linear inseparability, SVM introduces slack variables $ξ_{i} \geq 0, i = 1, \dots, l$ , which allow misclassification error with small samples and convert hard interval classification into soft margin classification. The optimization problem of classification hyperplane is given as follows

{\begin{matrix} min_{w, b, ξ} \frac{1}{2} ‖ w ‖^{2} + C \sum_{i = 1}^{l} ξ_{i} \\ s . t . y_{i} (w^{T} \cdot x_{i} + b) \geq 1 - ξ_{i}; ξ_{i} \geq 0; i = 1, 2, \dots, l \end{matrix}

(3)

Here, C is the classification error penalty factor. The problem of linear constrained quadratic programming is solved using Lagrange’s optimization method, that is

\begin{matrix} \max_{α_{i} β} min_{ω \cdot b \cdot ξ} {L_{p} = \frac{1}{2} {‖ ω ‖}^{2} + C \sum_{i = 1}^{-} ξ_{i} - \sum_{i = 1}^{l} α_{i} [y_{i} (ω \cdot x_{i} + b) - 1 + ξ_{i}] - \sum_{i = 1}^{l} β_{i} ξ_{i}} \\ s . t . α_{i} \geq 0, β_{i} \geq 0 \end{matrix}

(4)

Here, $α_{i}$ and $β_{i}$ are Lagrange multipliers. We, respectively, solve the partial derivatives of formula (4) with respect to $ω$ , $ξ$ , and b by setting the partial derivatives to zero. Then, the formula of solving dual optimization problem is denoted as follows

{\begin{matrix} \max_{α} {L_{D} = \sum_{i = 1}^{l} α_{i} - \frac{1}{2} \sum_{i . j = 1}^{l} y_{i} y_{j} α_{i} α_{j} (x_{i} \cdot x_{j})} \\ s . t . 0 \leq α_{i} \leq C, \sum_{i = 1}^{l} α_{i} y_{i} = 0 \end{matrix}

(5)

$α$ can be solved using the equation (5). Among them, $x_{i}$ corresponding to non-zero $α_{i}$ are support vectors of classifier. Then, optimal hyperplane $ω$ and b can be obtained according to the support vectors.

In dealing with the nonlinear classification problem, the input samples of SVM are mapped from the original space to a high-dimensional feature space through nonlinear mapping function $ϕ$ , in which the classification hyperplane $w \cdot ϕ (x) + b = 0$ is established and the optimal classification function is denoted by $f (x) = sgn {\sum_{i = 1}^{l} y_{i} α_{i}^{*} [Φ (x_{i}) \cdot Φ (x)] + b^{*}}$ . Based on the functional theory, in the optimal classification plane, the nonlinear classification is converted to the linear classification by the inner product function $K (x, y)$ which satisfies Mercer’s condition. Different algorithms can be obtained by selecting different inner product kernel functions.

OCSVM

In the process of surface defect classification, sometimes some new types of defects occur and need to be classified, but the number of defect samples of the new types is small and is not balanced compared with the number of training defect samples of other types. In this case, the normal SVM classification method will reduce the classification accuracy.

Along with thorough research and an urgent need for engineering application, some new algorithms about SVM are constantly found out, such as least squares support vector machine (LS-SVM),²² fuzzy SVM,²³ OCSVM,^24,25 and multi-class support vector machine (MSVM).^26,27

OCSVM was originally proposed for outlier detection, which was regarded as one-class classification for solving the quadratic programming problem. OCSVM has effectively solved the problem of high dimension, local minima, and nonlinear classification. Tax and colleagues^25,28,29 proposed a support vector data description (SVDD) classifier, which was realized by constructing the hypersphere method. B Schölkopf et al.^24,30 presented OCSVM classification algorithm, which was realized by constructing the hyperplane separated from the origin. OCSVM has been widely used in fault diagnosis, network intrusion detection, target recognition, and so on.^31–33

The basic idea of OCSVM is to describe the objects as a whole, to construct a closed and compact area, and to make the described object be included in the region as much as possible. Given the training sample points sets ${x_{i}, i = 1, 2, \dots, N}, x_{i} \in R^{N}$ , assuming that there is nonlinear mapping $ϕ$ from $R^{N}$ to a high-dimensional feature space $χ$ , that is, $ϕ (x_{i}) \in χ$ . A hypersphere (center a and radius R) with minimum volume is found to enclose $x_{i}$ as much as possible. To reduce the impact of singular points, slack variable $ξ_{i}$ is introduced and allowed errors to some extent. The optimization problem is given as follows

min [R^{2} + C \sum_{i} ξ_{i}], ξ_{i} \geq 0

(6)

where C is penalty factor which plays the role of controlling the punishment degree of error-classifying samples to achieve a compromise between the proportional extent error-classifying samples and the complexity of algorithms. According to the Lagrange’s multiplier method

\begin{array}{l} L (R, a, ξ_{i}) = R^{2} + C \sum_{i} ξ_{i} \\ - \sum_{i} λ_{i} {R^{2} + ξ_{i} - (x_{i}^{T} x_{i} - 2 a^{T} x_{i} + a^{T} a)} - \sum_{i} γ_{i} ξ_{i} \end{array}

(7)

Here, $λ_{i} \geq 0$ and $γ_{i} \geq 0$ are Lagrange coefficients. Solving the minimum value of equation (7) can be transformed to solve the maximum value of the dual problem as follows

L (λ_{i}) = \sum_{i} λ_{i} (x_{i} \cdot x_{i}) - \sum_{i, j} λ_{i} λ_{j} (x_{i} \cdot x_{j})

(8)

A new sample of z is determined whether it belongs to the target sample by the following formula

\begin{array}{l} (z - a) (z - a)^{T} = (z \cdot z) - 2 \sum_{i} a_{i} (z \cdot x_{i}) \\ + \sum_{i, j} a_{i} a_{j} (x_{i} \cdot x_{j}) \leq R^{2} \end{array}

(9)

If the formula (9) is satisfied, it is determined that z belongs to the target sample, otherwise it is judged to be a non-target sample and refused. Radius R can be determined by calculating the distance z to a support vector $x_{i}$ on the boundary. During numerical modeling, most values of $a_{i}$ are 0, and only a few aren’t 0. These samples corresponding to $a_{i} \neq 0$ are called support vectors, which determine the values of a and R. Other non-support vectors correspond with $a_{i} = 0$ , so it can be ignored in computing.

Kernel parameter optimization based on particle swarm algorithm

Kernel functions, as the key of SVM method, mainly contain the polynomial kernel function, radial basis function (RBF) kernel function, and sigmoid kernel function. The selection of kernel function does not have determined method and specific theoretical basis. Vapnik et al.³⁴ have proved that in the above three kinds of kernel functions, regardless of the sample size or the dimension level, RBF kernel functions are all applicable.

RBF kernel function has strong self-learning and self-regulation ability, and it is an ideal classification function. Therefore, this study employs an RBF kernel function in the SVM to classify TFT-LCD surface defect samples. The RBF is as follows

K (x, x_{i}) = \exp (- g {‖ x - x_{i} ‖}^{2})

(10)

OCSVM is modeled based on RBF kernel function. The model has two main parameters: the penalty factor C and the kernel parameter g. To get a good SVM classifier, the key is to select and optimize the value of nuclear parameters C and g. Parameter C is used to achieve a compromise between model complexity and misclassification ratio. If the C value is too large, it will lead to the reduction in SVM generalization ability. On the contrary, if the value is too small, it would affect the classification results. Parameter g represents the width of the kernel function, which plays an important role in improving the performance of SVM classification.

The common parameter optimization algorithms^35–38 include genetic algorithm (GA), grid search algorithm (GSA), and so on. These algorithms when dealing with high-dimensional complex problems will easily fall into local extreme accuracy and low defects. However, when the algorithms are used to deal with complex high-dimensional optimization problems, they can easily cause the problems of local extrema and low accuracy. To solve this problem, PSO algorithm is selected to optimize and determine the parameters.

PSO, developed by Professors Kennedy and Eberhart,³⁹ is an efficient parallel optimization algorithm based on swarm intelligence, which is originated from the predatory behavioral research of birds. PSO is a global search algorithm based on iterative update, which searches for the optimal solution through collaboration and information sharing among and between individuals.⁴⁰ Because of the advantages of parameter-less and fast convergence, it has been widely applied in function optimization, pattern recognition, and so on.^41–43 Its principle can be described below.

Assume that the population consists of m particles. A particle is seen as a point in D dimensional space, which represents the solution of the problem and its status is described by its position and speed. In D dimensional space, the position vector and velocity vector of the $i th$ particle are, respectively, expressed as $X_{i} = [x_{i 1}, x_{i 2}, \dots, x_{iD}]$ and $V_{i} = [v_{i 1}, v_{i 2}, \dots, v_{iD}]$ .The best individual historical position of particles is $P_{g} = [p_{g 1}, p_{g 2}, \dots, p_{gD}]$ , and the neighborhood best position of particle i is $P_{i} = [p_{i 1}, p_{i 2}, \dots, p_{iD}]$ . Particles constantly update their velocity and position in order to find the optimal solution.

The velocity and position of each particle are updated by formula (11)

{\begin{matrix} V_{i}^{t + 1} = ω V_{i}^{t} + c_{1} r_{i}^{t} (P_{i}^{t} - X_{i}^{t}) + c_{2} r_{2}^{t} (P_{g}^{t} - X_{i}^{t}) \\ X_{i}^{k + 1} = X_{i}^{k} + V_{i}^{k} \end{matrix}

(11)

Here, $ω$ is the inertia weight and k is the iteration number. $r_{1}$ and $r_{2}$ are uniformly distributed random numbers in the interval (0,1). $c_{1}$ and $c_{2}$ are the learning factors of $P_{i}$ and $P_{g}$ , which are usually between [0, 2]. The velocity of every particle is limited to $[- V_{\min}, V_{\max}]$ in order to avoid missing the optimal solution. The number of iterations to achieve the optimal solution meeting the demand is then determined and is used as the termination condition of the algorithm.

Penalty parameter C and kernel parameter g are optimized by PSO and OCSVM. The flowchart of the algorithm is shown in Figure 6:

Step 1. Set up the search range of SVM model and PSO parameters (population size, learning factors, inertia weight, maximum number of iterations, and maximum speed).

Step 2. The values of C and g generated randomly are used as the initial position of each particle. Initialize the velocity and position of all particles randomly.

Step 3. Initialize the individual optimal and global optimal locations of particles and calculate individual fitness value of the population. Compare individual fitness value with local optimal solution and global optimal solution. According to the comparison results, update local optimal solution and global optimal solution.

Step 4. Update velocity and position of all particles according to the results of step 3.

Step 5. If the maximum number of iterations is reached, return the optimal solution and terminate optimization; otherwise, continue to implement Step 3.

Figure 6.

The flowchart of optimization parameters of PSO-OCSVM algorithm.

Experimental results and analysis

In total, 60 samples are prepared for each type of defect images (pinholes, scratches, and particles). For each type, 40 samples are extracted as the training set by random sample generation method, and the rest 20 samples are used as the test sample set.

For OCSVM classification model, the search range of penalty factor C is [0.0001, 100]; and the search range of kernel parameter g is [0.0001, 100]. For PSO algorithm, the learning factor is c₁ = c₂ = 2, the maximum number of iterations is 50, the number of particles dimension is 2, the maximum velocity is 1, the maximum inertia weight is 0.9, and the minimum inertia weight is 0.3.

In order to evaluate the performance of the classification method, GA, GSA, and PSO are also used to optimize the SVM model parameters: kernel parameter g and penalty parameter C. Five principal component characteristic parameters are selected as the model input parameters. Table 1 is the classification results of TFT-LCD micro defect test samples based on these algorithms. The following can be drawn from the table:

The highest accuracy rate is 91.7%, which caused mainly by two reasons. On the one hand, the classification accuracy rate of micro-defects is relatively lower than macro defects’. On the other hand, the illumination has a certain influence on the feature extraction of defect area and causes some false positives and false negatives, such as broken line and hole shadow.

The performance of parameter optimization algorithm is better than the normal algorithm (kernel parameter g and penalty parameter C are randomly selected). The accurate recognition rate of normal algorithm is only 80%. PSO algorithm is proved to get the optimal model parameter (C = 64.637 and g = 0.359), and the recognition accuracy of 91.7% is obtained.

PSO-OCSVM with excellent performance and high reliability can be applied to the classification of TFT-LCD surface micro-defects.

Table 1.

Results of OCSVM model based on different parameter optimization methods.

Defect type	Normal algorithm		PSO		GA		Grid search
Defect type	Correct identification	Correct rate (%)	Correct identification	Correct rate (%)	Correct identification	Correct rate (%)	Correct identification	Correct rate (%)
Pinholes	16	80	18	90	18	90	17	85
Scratches	15	75	18	90	16	80	16	80
Particles	17	85	19	95	18	90	18	90
Results	48	80	55	91.7	52	86.7	51	85

OCSVM: one-class support vector machine; PSO: particle swarm optimization; GA: genetic algorithm.

Conclusion

The rapid detection method of TFT-LCD surface micro-defects is proposed based on PSO and OCSVM. An image acquisition system is constructed to acquire the surface micro-defects images of TFT-LCD. Background textures are removed by 1D DFT. WT algorithm is used to eliminate the influence of uneven illumination. Effective characteristic parameters describing TFT-LCD surface micro-defects are selected by PCA. Prediction model is developed based on OCSVM using RBF. By comparing with the algorithms of normal, GA, GSA, and PSO, PSO algorithm is proved to get the optimal model parameter (C = 64.637 and g = 0.359), and the recognition accuracy of 91.7% is obtained. The results indicate that the proposed TFT-LCD surface defect detection system and the PSO-OCSVM method can be used to inspect micro-defects with high classification accuracy.

Footnotes

Academic Editor: Gang Chen

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported by a grant from the National Key Scientific Instrument and Equipment Development Projects, China (2013YQ220749), Key Project of Science and Technology Research of Henan Province Education Department, China (16A510028), the National Natural Science Foundation of China (U1304305).

References

Jian

CX.

Review of TFT-LCD surface defect detection methods. Video Eng 2015; 39: 146–152.

Shi

AOI techniques for surface defect inspection. Appl Mech Mater 2010; 36: 297–302.

Shindo

Wang

Akella

. Effective excursion detection by defect type grouping in in-line inspection and classification. IEEE T Semiconduct M 1999; 12: 3–10.

Nakashima

. Hybrid inspection system for LCD color filter panels. In: Proceedings of the 10th international conference on instrumentation and measurement technology, Hamamatsu, Japan, 10–12 May 1994, pp.689–692. New York: IEEE.

Taniguchi

Ueta

Tatsumi

A mura detection method. Pattern Recogn 2006; 39: 1044–1052.

Vapnik

VN.

Statistical learning theory. New York: John Wiley & Sons, 1998.

Burges

CJC

. A tutorial on support vector machines for pattern recognition. Data Min Knowl Disc 1998; 2: 121–167.

Schölkopf

Smola

Learning with kernels. Cambridge, MA: MIT Press, 2001.

Smola

Schölkopf

A tutorial on support vector regression. Stat Comput 2004; 14: 199–222.

10.

Yan

Yen

ZC.

High-speed inline defect detection for TFT-LCD array process using a novel support vector data description. Expert Syst Appl 2011; 38: 6222–6231.

11.

Chi

Yung

In-TFT-array-process micro defect inspection using nonlinear principal component analysis. Int J Mol Sci 2009; 10: 4498–4514.

12.

Szu

. Automatic target defect identification for TFT-LCD array process inspection using kernel FCM-based fuzzy SVDD ensemble. Expert Syst Appl 2009; 36: 1978–1998.

13.

Yan

Yen

JC.

Fast support vector data descriptions for novelty detection. Neural Networks 2010; 21: 1296–1313.

14.

Tsai

Hung

CY.

Automatic defect inspection of patterned TFT-LCD panels using 1-D Fourier reconstruction and wavelet decomposition. Int J Prod Res 2005; 43: 4589–4460.

15.

Malik

Humayun

Parallelization of discrete cosine transform in shape from focus. Adv Sci Lett 2013; 19: 1459–1462.

16.

Ahmed

Natarajan

Rao

KR.

Discrete cosine transform. IEEE T Comput 1974; 100: 90–93.

17.

Mallat

SG.

A theory for multi-resolution signal decomposition: the wavelet representation. IEEE T Pattern Anal 1989; 11: 674–693.

18.

Gabor

Theory of communication. J Inst Electric Eng Jpn 1946; 93: 429–457.

19.

Vapnik

The nature of statistical learning theory. New York: Springer-Verlag, 1995.

20.

Yao

Chen

Cao

. Short-term traffic speed prediction for an urban corridor. Comput-Aided Civ Inf 2016; 31: 1–16.

21.

Song

Guan

. k-nearest neighbor model for multiple-time-step prediction of short-term traffic condition. J Transp Eng: ASCE 2016; 142: 04016018– 104016018-10.

22.

Suykens

VandeWalle

Least square support vector machine classifiers. Neural Process Lett 1999; 9: 293–300.

23.

Lin

Wang

SD.

Fuzzy support vector machines. IEEE T Neural Networ 2002; 13: 464–471.

24.

Schölkopf

Williamson

Smola

. Support vector method for novel detection. Adv Neur In 1999; 12: 526–532.

25.

Tax

DMJ

Pam

Duin

RPW.

Support vector data description applied to machine vibration analysis. Mach Learn 2000; 54: 45–66.

26.

Wang

Yao

. A comparison of the performance of ANN and SVM for the prediction of traffic accident duration. Neural Netw World 2016; 26: 271–287.

27.

Hsu

Lin

CJ.

A comparison of methods for multi-class support vector machines. IEEE T Neural Networ 2002; 13: 415–425.

28.

Tax

DMJ

Duin

RPW

. Support vector domain description. Pattern Recogn Lett 1999; 20: 1191–1199.

29.

Tax

DMJ

Duin

RPW

. Data domain description using support vectors. In: Proceedings of the European symposium on artificial neural networks, Bruges, 21–23 April 1999, pp.251–256. D-Facto.

30.

Schölkopf

Estimating the support of a high-dimensional distribution. Haifa: Department of Computer Science, University of Haifa, 2001.

31.

Wang

Wrong

Miner

Anomaly intrusion detection using one class SVM. In: Proceedings of the 5th annual IEEE SMC information assurance workshop, West Point, NY, 10–11 June 2004, pp.358–364. New York: IEEE.

32.

Wang

Yeung

Tsang

ECC

. Structured one-class classification. IEEE T Syst Man Cy B 2006; 36: 1283–1294.

33.

Shahid

Naqvi

Qaisar

. One-class support vector machines: analysis of outlier detection for wireless sensor networks in harsh environments. Artif Intell Rev 2015; 43: 515–563.

34.

Vapnik

Golowich

Smola

Support vector method for function approximation, regression estimation, and signal processing. In: Mozer

Jordan

Petsche

(eds) Advances in neural information processing systems 9. Cambridge, MA: MIT Press, 1997, pp.281–287.

35.

Peng

Shan

Guan

. Stable vessel-cargo matching in dry bulk shipping market with price game mechanism. Transport Res E: Log 2016; 95: 76–94.

36.

Zhu

Cai

. Two-phase optimization approach to transit hub location—the case of Dalian. J Transp Geogr 2013; 33: 62–71.

37.

Yao

Gao

. An improved particle swarm optimization for carton heterogeneous vehicle routing problem with a collection depot. Ann Oper Res 2016; 242: 303–320.

38.

Kong

Sun

. A bi-level programming for bus lane network design. Transport Res C: Emer 2015; 2015: 310–327.

39.

Kennedy

Eberhart

Particle swarm optimization. In: Proceedings of IEEE international conference on neural networks, vol 4, Piscataway, NJ, 27 November–1 December 1995, pp.1942–1948. New York: IEEE.

40.

Ranaee

Ebrahimzadeh

Ghaderi

Application of the PSO-SVM model for recognition of control chart patterns. ISA T 2010; 49: 577–586.

41.

Zhang

Berg

Maire

. SVM-KNN: discriminative nearest neighbor classification for visual category recognition. In: Proceedings of the IEEE computer society conference on computer vision and pattern recognition, New York, 17–22 June 2006, pp.2126–2136. New York: IEEE.

42.

Vapnik

Chervonenkis

The necessary and sufficient conditions for consistency in the empirical risk minimization method. Pattern Recogn Imag Anal 1991; 1: 283–305.

43.

Kapp

Sabourin

Maupin

. A PSO-based framework for dynamic SVM model selection. In: Proceedings of the genetic and evolutionary computation conference, Montreal, QC, Canada, 8–12 July 2009, pp.1227–1234. New York: ACM.