Visual defects sometimes occur during the manufacturing of flat panel liquid crystal displays (LCDs). One class of defects includes a variety of blemishes variously called stain (English), mura (Japanese) or alluk (Korean). These blemishes appear as low contrast, non-uniform brightness regions, typically larger than single pixels. They are caused by a variety of factors such as non-uniformity distributed liquid crystal material and foreign particles within the panel. Such blemishes cannot be repaired. Automatic inspection systems, designed for pixel and line defect detection, have had difficulty accurately detecting and quantifying LCD blemishes. At present, most blemish detection is performed by human inspectors. This paper describes a recently developed automatic inspection system, which reliably detects, quantifies and classifies LCD blemishes in the presence of single pixel and line pixel defects that tend to obscure the subtle blemishes. The algorithm underlying this system, called MuraLookTM, uses conventional image processing operators such as convolutional filtering, morphological filtering and blob shape analysis under region-of-interest control in a novel combination to systematically separate each of over twenty different blemish patterns. Strength measures for each class of blemish are used under human operator control to grade each blemish as pass or fail. The paper discusses various types of defects in LCD panels and relates them to the MuraLook system defect class patterns. The architecture of the MuraLook defect detection system is described.
We have developed an automated visual inspection technology for hard disk drive head suspensions. It consists of foreign materials detection technique and a precise, high-speed pattern width measurement. High S/N-ratio sensing is necessary for foreign materials detection. We developed a spectral sensing method that senses the reflection of two colors, blue and red, and calculates the reflection coefficients of the colors for every pixel. With this method, the S/N-ratio is enhanced by seven times. A high accuracy is required for measuring pattern widths. The problem is that this must be achieved within an inspection time of 0.4 s/suspension, which conventionally takes several seconds. To obtain a pattern signal at high speeds, suspensions are illuminated with transmitted light and their images are captured with a CCD line sensor. A sub-pixel method that calculates interstitial signal level enables both a high accuracy and high-speed measurement. An automated optical inspection system utilizing these technologies is currently operating in one of Fujitsu's factories. This paper describes this optical system, its measuring procedures, and the measured results of the system.
The laser cladding process consists in adding a molten blown powder to a partially melted substrate, this in order to change the substrate properties. The aim of this paper is to present different experimental set-up associated with different image processes in order to characterize the powder stream in the laser cladding process. Information such as the particle speed, the powder distribution in function of the nozzle distance are presented. The used algorithms as well as the experimental set-up are presented and detailed.
Knowledge of internal log defects, obtained by scanning, is critical to efficiency improvements for future hardwood sawmills. Nevertheless, before computed tomography (CT) scanning can be applied in industrial operations, we need to automatically interpret scan information so that it can provide the saw operator with the information necessary to make proper sawing decisions. Our current approach to automatically label features in CT images of hardwood logs classifies each pixel individually using a back-propagation artificial neural network (ANN) and feature vectors that include a small, local neighborhood of pixels and the distance of the target pixel to the center of the log. Initially, this ANN was able to classify clearwood, bark, decay, knots, and voids in CT images of two species of oak with 95% pixel-wise accuracy. Recently we have investigated other ANN classifiers, comparing 2D versus 3D neighborhoods and species-dependent (single species) versus species- independent (multiple species) classifiers using oak, yellow poplar, and cherry CT images. When considered individually, the resulting species-dependent classifiers yield similar levels of accuracy (96 - 98%). 3D neighborhoods work better for multiple-species classifiers and 2D is better for single-species. Under certain conditions there is no statistical difference in accuracy between single- and multiple-species classifiers, suggesting that a multiple- species classifier can be applied broadly with high accuracy.
In joining defects on semiconductor wafer maps into clusters, it is common for defects caused by different sources to overlap. Simple morphological image processing tends to either join too many unrelated defects together or not enough together. Expert semiconductor fabrication engineers have demonstrated that they can easily group clusters of defects from a common manufacturing problem source into a single signature. Capturing this thought process is ideally suited for fuzzy logic. A system of rules was developed to join disconnected clusters based on properties such as elongation, orientation, and distance. The clusters are evaluated on a pair-wise basis using the fuzzy rules and are joined or not joined based on a defuzzification and threshold. The system continuously re- evaluates the clusters under consideration as their fuzzy memberships change with each joining action. The fuzzy membership functions for each pair-wise feature, the techniques used to measure the features, and methods for improving the speed of the system are all developed. Examples of the process are shown using real-world semiconductor wafer maps obtained from chip manufacturers. The algorithm is utilized in the Spatial Signature Analyzer software, a joint development project between Oak Ridge National Lab. and SEMATECH.
One of the main problems in machine vision inspection systems which use the triangulation of structured light is improving the accuracy of the laser light centroid algorithm. This paper presents a new and practical way to calculate the center of reflection of structured laser light. Reflection from a laser light line has a normal distribution of light intensity. Traditional algorithms ignore the influence of specular reflecting light and assume the laser light line intensity is not a normal distribution. The new algorithm considers this situation and gives more weight to the central sampling values when calculating the center of the reflected laserlight. Experiments verify that this algorithm is more robust and accurate than traditional algorithms.
We developed a high-speed 3D inspection system for solder bumps. The system applies laser-based triangulation with a laser diode, an acousto-optical deflector (AOD), and a position sensitive detector. The system scans LSI surfaces with a laser beam at a 15 m/s scanning speed, and can acquire height data at rate of up to 7 X 106 samples/sec with 0.8 micrometers resolution. For high-speed laser beam scanning with the AOD, we developed a technique to correct for the cylindrical lensing effect that causes astigmatism on a focal plane. Our correction method is unique in that it notices the scanning speed being constant in order for the dynamic deviation to be erased. This technique can suppress these deviations enough to enable accurate laser scanning. When installed on an LSI chip assembly line, the system can measure each bump height to within +/- 2 micrometers in 5 ms. This will allow for a 100% inspection to be achieved.
In many web inspection applications the inspection should be able to detect and classify a large number of different- sized defects with varying scattering properties. As a consequence, a high-resolution system with a wide dynamic range is needed. The performance of the system should also remain uniform over the image area. Three major elements affecting the image formation of a web inspection system are illumination, imaging and detection algorithms. The relationship between these elements and the final image quality is discussed. Practical examples of how the system performance is related to the quality of the image formation are given. In the examples small or low-contrast defect samples picked from industrial manufacturing process are analyzed. Defects are classified as small if they cover an area of ten CCD pixels or less in the image plane or they have such an orientation that the size of the defect in one dimension is extremely small as is in the case of some scratches. The defect is considered as low-contrast if the relative defect contrast is less than the pattern noise in the imaging system. As a conclusion some criteria and an approach for the systematization of the design of the image formation are discussed.
Karnal bunt is a fungal disease which infects wheat and, when present in wheat crops, yields it unsatisfactory for human consumption. Due to the fact that Karnal bunt (KB) is difficult to detect in the field, samples are taken to laboratories where technicians use microscopes and methodically search for KB teliospores. AlliedSignal Federal Manufacturing & Technologies, working with the Kansas Department of Agriculture, created a system which utilizes pattern recognition, feature extraction, and neural networks to prototype an automated detection system for identifying KB teliospores. System hardware consists of a biological compound microscope, motorized stage, CCD camera, frame grabber, and a PC. Integration of the system hardware with custom software comprises the machine vision system. Fundamental processing steps involve capturing an image from the slide, while concurrently processing the previous image. Features extracted from the acquired imagery are then processed by a neural network classifier which has been trained to recognize `spore-like' objects. Images with `spore-like' objects are reviewed by trained technicians. Benefits of this system include: (1) reduction of the overall cycle-time; (2) utilization of technicians for intelligent decision making (vs. manual searching); (3) a regulatory standard which is quantifiable and repeatable; (4) guaranteed 100% coverage of the cover slip; and (5) significantly enhanced detection accuracy.
A template-based vision system for the 100% inspection of printed flaws on green ceramic tape has been developed. Design goals included a requirement for the detection of flaws as small as two thousandths of an inch on parts up to 8 by 8 inches in size. The inspection engine is a Datacube, Inc., MV200 pipeline processor. As each part is inspected, four 2K by 2K pixel quadrant images are stitched together to construct a single 4K by 4K pixel image with the aid of multiple fiducials located in each quadrant. The part fiducial locations, mask image, and punched-hole position data are generated, beforehand, from CAD designs using a defect map editor (DME), a preprocessing software packaged developed for the PC. The DME also generates a part `defect map'. Each unique structure in the printed pattern is defined as an object. Objects are grouped into user-defined categories such as die pads, contact fingers, traces, and electrolysis buses. The map is used during the runtime inspection to associate each detected defect with an object group and a particular defect specification for that group. Repeat defects are optionally tracked for up to three consecutive parts.
We present a hybrid shape recognition system with an optical Hough transform processor. The features of the Hough space offer a separate cancellation of distortions caused by translations and rotations. Scale invariance is also provided by suitable normalization. The proposed system extends the capabilities of Hough transform based detection from only straight lines to areas bounded by edges. A very compact optical design is achieved by a microlens array processor accepting incoherent light as direct optical input and realizing the computationally expensive connections massively parallel. Our newly developed algorithm extracts rotation and translation invariant normalized patterns of bright spots on a 2D grid. A neural network classifier maps the 2D features via a nonlinear hidden layer onto the classification output vector. We propose initialization of the connection weights according to regions of activity specifically assigned to each neuron in the hidden layer using a competitive network. The presented system is designed for industry inspection applications. Presently we have demonstrated detection of six different machined parts in real-time. Our method yields very promising detection results of more than 96% correctly classified parts.
In this paper a High-speed Opto-electric Image Processing Unit (HSOEIPU) based on an incoherent correlator is constructed, which can do two main processing works. One is digital filtering; the other is object recognition. Using HSOEIPU a Lapalace filter is realized to extract the edge characteristic from the images. Then the preprocessed images are sent into the object recognition unit for identifying the important areas or targets. A novel template-matching method is proposed for gray-tone image recognition. A changeable cycle-encoding method is introduced to realize the fuzzy relation pixel-matching on a correlator structure. The system has good fault-tolerance ability for rotation distortion, gaussian noise disturbance or information losing. In the experiments photographs of workpieces are sent into HSOEIPU for processing.
An optoelectronic morphological processor for industrial on- line inspection is presented in this paper. The principle of the processor is based on the morphological hit-miss transform. By using an extensive complementary encoding method, which combines the foreground and background of an image into a whole, the hit-miss transform, which usually needs three steps to perform, can be implemented in only one step. The optical implementation hardware is based on an incoherent optical correlator due to its compactness in structure and immunity to coherent noise. The experimental results are given.
The use of supervised pattern recognition technologies for automation in the manufacturing environment requires the development of systems that are easy to train and use. In general, these systems attempt to emulate an inspection or measurement function typically performed by a manufacturing engineer or technician. This paper describes a self- optimizing classification system for automatic decision making in the manufacturing environment. This classification system identifies and labels unique distributions of product defects denoted as `signatures'. The technique relies on encapsulating human experience through a teaching method to emulate the human response to various manufacturing situations. This has been successfully accomplished through the adaptation and extension of a feature-based, fuzzy k- nearest neighbor (k-NN) defined classes so that a significant reduction in feature space and problem complexity can be achieved. This k-NN implementation makes extensive use of hold-one-out results and fuzzy ambiguity information to optimize its performance. A semiconductor manufacturing case study will be presented. The technique uses data collected from in-line optical inspection tools to interpret and rapidly identify characteristic signatures that are uniquely associated with the manufacturing process. The system then alerts engineers to probable yield-limiting conditions that require attention.
The major drawbacks of automated visual inspection systems are the high set-up costs resulting from hard- and software development costs, labor, and maintenance costs. The key to solving the problem of flexibility is the development of visual inspection systems which are able to inspect a large variety of different objects without or only partly changing the analysis algorithm. One aspect in this design is the representation of the object to be inspection. A priori knowledge is used implicitly or explicitly by all visual inspection systems, since inspection can only be performed by matching the object under inspection with a set of predefined conditions of acceptability. These specifications are described by an explicit object model, that includes all relevant object features. The second representation of the object is the image containing the object. Within the image, object features are represented as image features, that have to be detected by feature detection algorithms. This paper shows an inspection model that allows a flexible object- specific description, defined in a so-called description language. This model has primitives as nodes and relations between the primitives as arcs. Furthermore tolerances and weights indicating the importance of detection are also part of the model. On the case study of analogue display instruments the representation and generation of the inspection model is shown.
A non-contact laser sensor based on the circular optical cutting image is developed to measure the size and the profile of the pipe inner wall. The sensor consists of a laser diode light source, an optical ring pattern generator and a CCD camera. The circular light from the optical ring pattern generator projects onto the pipe inner wall, which is then viewed by the CCD camera. The adapt weighted average value filter and subpixel technique in several step computing and half Gauss fitting are put forward to obtain the edge and the center of the circular image in order to filter the noise of the image and raise the resolution of the measuring system. The experimental results show that the principle is correct and the techniques are realizable.