technisat tft lcd reinigungsset factory
(Yicai Global) Aug. 11 -- Electronic display giant BOE Technology Group Co. [SHE:000725] has signed an agreement with the municipal government of Wuhan, in China"s central Hubei province, to invest CNY46 billion (USD6.9 billion) to build a factory for 65-inch LCD TFT-LCD display modules and related projects in the city, BOE said in a press briefing on Aug. 10.
TFT displays are full color LCDs providing bright, vivid colors with the ability to show quick animations, complex graphics, and custom fonts with different touchscreen options. Available in industry standard sizes and resolutions. These displays come as standard, premium MVA, sunlight readable, or IPS display types with a variety of interface options including HDMI, SPI and LVDS. Our line of TFT modules include a custom PCB that support HDMI interface, audio support or HMI solutions with on-board FTDI Embedded Video Engine (EVE2).
According to real LCD manufacturing conditions, the number of normal LCD panels exceeds greatly the number of defective LCD panels. Therefore, the normal PRs greatly outnumber the defective PRs. As a result, the collected data set for training would be imbalanced if a two-class classification approach is adopted, the SVM by Vapnik [4] for example, the class imbalance problem occurs.
In practice, in addition to the class imbalance problem, the LCD defect detection also suffers from another critical problem resulting from the absence of negative information. To facilitate the following problem description, the normal PR class and the defective PR class are defined as the positive class and negative class, respectively.
The main difference between a normal PR and a defective PR is that their appearances are apparently different, as can be observed from Figure 4. The color (or gray level) of a normal PR is nearly uniform, implying that the variation of the gray-level distribution of normal PRs is very small. On the contrary, the surfaces of defective PR not only contain various kinds of textures, but also vary greatly in color, implying that the variation of the true distribution for negative class in the data space is very large. Collecting a set of positive training data that can represent the true distribution of positive class is easy, because: (1) the variation of positive-class distribution is very small; and (2) most of the LCD panels are normal (the number of normal PRs is considerably large). Therefore, the positive class can be well-sampled during the data collection stage in real practice. However, representative defective PRs are difficult to obtain in practice for several reasons. For example, there are numerous types of defects in array process, more than 10 types at least. However, not all the defects would occur frequently. Some of the defects seldom appear, for example the defect caused by abnormal photo-resist coating (APRC). The defect “APRC” seldom occurs, because equipment/process engineers maintain the coating machines periodically. Even so, the coating machines might still break down occasionally. As a result, the number of available images containing the APRC defects is quite limited. But, the APRC defect has a large variation in color and texture. Unfortunately, limited APRC examples cannot stand for all kinds of APRC defects. Therefore, the collected negative training data are most likely under-sampled. Here, the “under-sampled” means that the collected negative training set cannot represent the true negative-class distribution in the data space, which is the problem of absence of negative information. Due to this problem, numerous false positive (i.e., missing defects) will be produced if a two-class classification approach (e.g., a binary SVM) is applied to the LCD defect detection, which has been evidenced by the results reported in [7]. Compared with two-class classification approach, novelty detection approach is a better choice.
Novelty detection is one-class classification [10,35], which is to solve the conventional two-class classification problems where one of the two classes is under-sampled, or only the data of one single class can be available for training [5,6,9–11,35–40]. As analyzed above, for the LCD defect detection application, the normal PRs can be well-sampled, while the defective PRs are in general undersampled. Therefore, the LCD defect detection can be treated as a typical novelty detection problem. Accordingly, one-class classification is a better solution.
To summarize, it can be seen that the LCD defect detection suffers from two problems simultaneously: one is the class imbalance problem, and the other is the problem of the absence of negative information. For the first problem, there have been many sophisticated solutions, including sampling, cost-sensitive learning, SVM-based, and one-class learning approaches. However, the only solution to the second problem is the novelty detection approach (i.e., one-class classification approach). Therefore, one-class classification would be a more appropriate approach to the LCD defect detection application.
There are several approaches for one-class classification, such as density approach (e.g., Gaussian mixture model [5]), boundary approach (e.g., SVDD [9] and one-class SVM [40]), neural network approach [6,36], and reconstruction-based approach (e.g., the kernel principal component analysis for novelty detection [35]). It has been proven in [9] that when a Gaussian kernel is used, the SVDD proposed by Tax and Duin [9] is identical to the one-class SVM proposed by Schölkopf et al. [40]. This paper focuses on the SVDD since it has been applied to the same application in the works of [7] and [10], and has shown to be effective in detecting defective PRs. However, as discussed in Section 1, generalization performance of SVDD is limited. Therefore, the intent of this paper is on proposing a method to improve generalization performance of SVDD, and applying the improved SVDD to the LCD defect detection treated as a novelty detection problem. The improved SVDD is called quasiconformal kernel SVDD (QK-SVDD). Note that the QK-SVDD and SVDD are not two independent classifiers. To obtain QK-SVDD, one has to train an SVDD first, which will be introduced in Section 2.4. In the following part of the paper, we first introduce the defect detection scheme, and then derive the proposed method in details.