TFT-LCD is widely used in various device applications in the display market. Since the TFT-LCD industry is now in the phase of maturity, it is important to make a domestic production of competitive product such as polarized film. Therefore, the importance of a polarized film is on the rise in line with high standards of required specification as polarizing plate industry is growing. Since the polarized film used in TFT-LCD is vulnerable to dusts, bubbles, and scratches which may occur potential causes of defection. So far, defect inspection has been depended on the inspector’s judgement. Therefore, there has been various attempts to develop automatic inspection system for cost reduction and quality improvement. Automatic inspection system for a polarized film is composed of detecting defective candidate pixel, blob construction and feature extraction, defect classification, and recognition stages.
Defects in the polarized film are classified by the major defect that the difference of intensity with surroundings is high and the ultimate defect that the detection of defects is difficult due to the small difference in intensity. Under-detection in the defect inspection system may cause a serious effect on product quality. Therefore, defect is classified through blob analysis after over-detection. However, the result of over-detection can cause a burden of blob analysis stage so that it is very important to minimize over-detection.
This paper proposes a defect detection algorithm with a precise detection of an ultimate defect over minimizing over-detection. The proposed defect detection process is as follows.
First of all, the proposed algorithm estimates the distribution of background region excluding defect region by using skewness of histogram at the defect detection stage. It estimates the intensity range of background region that minimizes image’s skewness based on the analysis of normal distribution histogram of image without having defect. Then this algorithm detects the candidate pixels and regions that are excluded from estimated background distribution.
Second stage detects additional candidate defect pixels based on a histogram distribution of local region in order to minimize noise of over-detection. If a local region shows heterogeneous characteristics, this algorithm decides the region as combined region of defect and background pixels. By using probability distribution estimation of heterogeneous region, the algorithm can detect additional candidates of defect pixels.
Finally, this algorithm calculates the probability density function based on Parzen window to classify detected defect candidate pixels between real defection pixel and noise. Finally defected pixel is selected by using the probability density function.
In the experiment, we used the simulation image that reflects the characteristics of real polarized film for a quantitative evaluation of minimizing over-detection and defect detection performance evaluation of the proposed method. Also, we applied the proposed algorithm using in a real polarized film image for highly reliable detection result to prove the performance of the proposed algorithm.