When modeling the background with kernel density estimation, the selection of a proper kernel bandwidth becomes a critical issue. It is not easy, however, to perform pixel-wise kernel bandwidth estimation when the data associated with each pixel is insufficient. In this paper, we present a new method using spatial information to estimate the pixel-wise kernel bandwidth. The number of pixels in a spatial region is large enough to capture the variance of the underlying distribution on which the optimal kernel bandwidth is estimated. To show the effectiveness of the estimated kernel bandwidth, the background subtraction using this bandwidth is applied to OLED defect detection and its result is compared to those using the bandwidths obtained from other approaches.
We present a new formulation to solve a defect detection problem on images using multiple reference images. The reference images are defect-free images obtained from the same position of other products. The defect detection problem is reformulated as a binary labeling problem, where each pixel is labeled with "one" if it contains a defect and with "zero" otherwise. The formulation of the energy function used for the labeling problem is defined. Then, the graph-cuts algorithm is used to obtain the optimal label set minimizing the energy function that becomes the defect detection result. The presented approaches are robust to noises taken from several sources, including image-taking, transmission process, environmental lighting, and pattern variation. It does not suffer from the alignment problem for the conventional comparison methods using references. These approaches are illustrated with real data sets, semiconductor wafer images collected by scanning electron microscope equipment, and compared to other defect detection approach.
One of widely used methods to extract boundaries of objects in the image is the contour method based on the energy
models. Two well known energy models are the edge based and the region based model. Although each of the two
models works well with its own advantage, it has difficulty in the following situation: When the initial point the contour
evolving starts from is not located properly in the inside region of an object or some objects are partially overlapped so
that the intensity difference between boundary pixels on the overlapped area and their neighbors becomes relatively
small, the edge based model approach fails to produce good results. On the other hand, the region based model approach
fails to produce good results when more than two objects with different intensity averages exist. To overcome such
difficulty, we suggest the hybrid energy model approach constructed partially from each of the two approaches. In this
approach, some initial point the contour evolving starts from is randomly selected from the image. From the selected
initial point the contour starts evolving until it meets the boundary of either some object or background. Once the
boundary of one object or background is found, its region is removed from the image. On the rest of the image, the same
procedure is repeated until the boundaries of all objects and background are found. The suggested approach is illustrated
using SEM images of semiconductor wafers.
We propose an evaluation system that assigns each line-type thin-film transistor liquid-crystal display defect a corresponding level that objectively agrees with human visual perception. By “objective,” we mean that the evaluation corresponds, on average, with the assessment of a group of inspectors. The basic idea is to use the human visual perception to evaluate defects. Crucial features of defects are selected to represent the human visual perception for the line-type defect. In the process, we define the “just-noticeable difference surface” (JND) and evaluate the level of defect as the distance from a feature point consisting of selected features of the JND.
Detecting defects is important technology of the TFT-LCD (Thin Film Transistor-Liquid Crystal Display)
production process for quality control. For high quality and improving productive, defect detection is performed on each
manufacturing process. In array process, defect inspection is divided into inspection for active matrix area and inspection
for pad area. Inspection on active matrix area has used period of pattern to detect defect. As pad has non repetitive
pattern, period can not be used for defect detection. Therefore, defects on pad have been detected by referential method
comparing to pre-stored reference pad image. Subtraction has been used for comparison with reference pad. This method
is problematic for pad defect inspection due to variance in the shapes of pad, illumination change and alignment error. In
this paper, we propose the inspection method making up for limitation of referential method which has been used for
TFT-LCD pad. Inspection is performed by applying morphological method to each horizontal line. By finding valley of
each line, defect is detected.
The thin film transistor liquid crystal display (TFT-LCD) has become an actively used front of panel display
technology with an increasing market. Intrinsically there is a region of non uniformity with low contrast that to human
eye is perceived as a defect. Because the grey-level difference between the defect and the background is small, the
conventional edge detection techniques are hardly applicable to detect these low contrast defects. Although several effort
were dedicated in classifying the patterned TFT-LCD defects, only few researches were conducted on detecting the unpatterned
TFT-LCD defects that accounts for approximately 15% of all defects produced during the manufacturing stages. This paper proposes a detection method for the un-patterned TFT-LCD defects by using the directional filter bank (DFB), Shen-Castan filter and maximum Feret's diameter. The effectiveness of the proposed method is tested through
the experiment using real TFT-LCD panel images.
Recently, with an increasing FPD market, automatic detection of the mura in the manufacturing process has become a critical issue for manufactures interested in increasing their TFT-LCD quality. But segmentation based detection algorithms deviate from human visual perception model. To supplement the detection error produced by deviation, the mura is re-inspected through a visual inspection during manufacturing process. If we could objectively quantify each mura's defect degree, then based on some threshold of defect degree, we could reduce the number of re-inspection. We call this degree line muras defect level. Our approach is an attempt to quantify the ideal defect level of line mura, that for each individual could vary because of subjectivity, based on multiple features crucial in the detection of line mura. In the process, we approximated what we call JND surface that passes through the middle of feature points with mean
mura visibility of 0.5. Then Index function, which measures distance from JND surface, is employed to measure the objective defect level of each candidate mura.
TFT-LCD generally has the intrinsic non-uniformity due to the variance of the backlight. The region that has the perceptible non-uniformity is defined as a defect, called area-mura. In this paper, we present a new segmentation method for detecting area-mura. We first extract candidates of area-muras using regression diagnostics and then select the real area-muras among those candidates based on the size and SEMU index, a measure of contrast based on human brightness perception. Performance of the presented method has been evaluated on those TFT-LCD panel samples provided by Samsung Electronics Co., Ltd.
Content of an image can be expressed in terms of different features such as color, texture, shape, or text annotations. Retrieval based on these features can be various by the way how to combine the feature values. Most of the existing approaches assume a linear relationship between different features, and also require the user to directly assign weights to features. In particular, as the number of feature classes increases, intuition about how to pick relative weights among features is lost. While this linear combining approach establishes the basis of content-based image retrieval (CBIR), the usefulness of such systems was limited due to the difficulty in representing human perception subjectivity. In this paper, we introduce a Neural Network- based Image Retrieval system, a human-computer interaction approach to CBIR using Radial Basis Function network. This approach determines nonlinear relationship between features so that more accurate similarity comparison between images can be supported and allows the user to submit a coarse initial query and continuously refine his information need via relevance feedback. The experimental results show that the proposed approach has the superior retrieval performance than the existing approaches such as linearly combining approach, the rank-based method, and the BackPropagation- based method.
In content-based image retrieval (CBIR), retrieval based on different features can be various by the way how to combine the feature values. Most of the existing approaches assume a linear relationship between different features, and the usefulness of such systems was limited due to the difficulty in representing high-level concepts using low-level features. In this paper, we introduce Neural Network-based Image retrieval (NNIR) system, a human-computer interaction approach to CBIR. By using the Radial Basis Function (RBF) network, this approach determines nonlinear relationship between features so that more accurate similarity comparison between images can be supported. The experimental results show that the proposed approach has the superior retrieval performance than the existing linear combining approach, the rank-based method and the Back Propagatoin-based method. Although the proposed retrieval model is for CBIR, it can easily be expanded to handle other media types such as video and audio.