Machine vision is widely used in die bonders to provide perfect alignment between the wafer or leadframe and the die pickup tip during pick and place operations respectively. This also results in better detection of bad ink-dotted dies and accurate placing of epoxy resin on the die bond pads of the leadframe. Pattern matching using Pattern Recognition techniques and area alignment using edge detection techniques are the two most widely used methods in providing die and leadframe pad alignment. The increased emphasis on quality improvement on semiconductor assembly has results in the use of machine vision for quality inspection in addition to alignment. The inspection is usually, done at two points in the die bond cycle, one just after epoxy application on the leadframe pad (pre-bond inspection) and the other after the die is bonded to the leadframe pad. The post bond inspect confirms if the placement of the die on the leadframe pad is within the required specification. It also looks for angular die placement as well as for epoxy spills on top of the die.

The mechanical integrity of the wire-bond connections in an integrated circuit governs the electrical integrity of the IC chip. Good quality control of bonding operation requires the wire-bond ball size on the aluminum pad of the silicon chip to be within certain specified range of diameters and heights. This paper describes the problems associated with the inspection of wire bonding balls height using 2D images. Two methods of inspecting the bonding ball height were described. The first method uses structure lighting approach to gauge the height and the second method estimates the height by computing the focus indices of a series of the 2D wire-bond images. Fourier energy spectral density of the gradient image is used as a basis for computing the focus index of the image. A simple and effective differential method was used to compute the gradient image. Both methods are able to achieve an accuracy of +/- 1 micrometers for wire- bond ball height measurement.

Machine vision is a well established method of automating mark inspection on IC packages. Defects such as missing mark, blurred mark, mis-orientation, illegible mark, etc. can be caught accurately and repeatedly by a good mark inspection system. The most common inspection algorithm that is employed is `correlation' (or template matching). Here the mark characters are taught to the system by the operator on one device. These characters are stored as a template set. Subsequent devices are compared with this stored template set and based on the comparison results the mark is classified as good or as reject. One of the setbacks of this traditional method of inspecting mark with a single template set is that it overkills when marking has variation. Variation in the mark can arise from the same or different marking machines and also when device lots with different mark characters are combined. Mark inspection using multiple templates provides a solution to this problem. In this method, instead of one template set, multiple template sets can be taught and subsequent devices are checked with these multiple mark template sets.

In pixel based setting, thinning is often accomplished by removing contour points in connected regions iteratively. Such algorithms require the image be presented in a regular spatial grid. Missing data points are not allowed. In range data processing, such requirements are not always fulfilled due to reflectance and sensor geometry. An algorithm is developed for extraction and thinning of 3D linear features. It is based on fitting of structuring element. The data is assumed to be 3D measurement points of the objects of interest. Three criteria are employed to extract and simplified the required linear feature. First, connected point sets are identified. For each connected set, an occurrence of linear structure is then located by fitting a cylindrical structuring element. Finally, the complete linear feature is identified by motion of the structuring element along the major axis. This algorithm is useful for 3D inspection.

We report on the development of an automated inspection and dimension measurement system for optoelectronic components. Images of the components are captured with a CCD camera. Precision templates are designed for image calibration, and a sub-pixel detection software is developed for dimension measurement. Experimental results demonstrate that the system is capable of measuring the dimensions with an accuracy of 30 micrometers .

Automatic inspection systems for IC mark, package and lead inspection are being widely used as in-process controls and check points. Here their primary function is not only to inspect and sort out defective parts but also to provide feedback on how well a process such as marking or trim and form is performing. Inspection results of every part inspected are often accumulated in a statistical process control (SPC) program that can monitor drifts in the process. Not all drifts are caused by problems in the process itself. For example the mark contrast on a package may be reduced not only because of some problem with the marking process but also because of changes in the mold compound of the package or changes in the light intensity of the inspection system. In latter case a statistical tool such as the SPC program may alert the user of a process drift and he will have to retune, recalibrate or change the parameters of the inspection system. Often the change in parameter is done by trail-and-error. A change too much or too little can result in excess overkill or even escapes. Alternatively the statistical data itself can be used to suggest the user what changes should be made to the inspection parameters. This method of automatic parameter optimization is discussed in detail in this paper. A mark inspection system is chosen as a specific example on how to apply this method.

This paper presents a system for automatic recognition of Integrated Circuit chips on assembled circuit boards by recognizing characters on its package label. The paper concentrates on the extraction and recognition of these character. The extraction of characters involves binarization of the image, character localization, orientation alignment and character re-sizing. A pre-trained two-layer feed-forward neural network is used to recognize the extracted characters. The neural network is trained using the back-propagation method. Experimental results are presented.

Texture is one of the important characteristics used in identifying objects or regions of interest in an image. Statistical approach algorithms for image classifications are very poor techniques in identifying texture in particular the spatial gray level dependence method (SGLDM). The main disadvantage is the intensive computation required for this algorithm. The advantage of using ANN is less computational time once the network is trained and constructed in a parallel architecture. To improve the computational speed and parallelism further a structured ANN is used. Here, we will describe the use of this ANN for textured image segmentation. A structural artificial neural network with three sub-networks is proposed to estimate the SGLDM algorithm. A texture image segmentation system can be built by using this network and searching window method. The advantage of this design is that the ANN structure is a feed-forward network, so that the system can be built in a pile-line fashion. One of the applications can be the object searching in wafer or VLSI circuit inspection.

During the stage of pre-cap inspection, 3D profiles of bonded wires are inspected to detect defects such as lifted, tight and sagging wires, and to measure the maximum wire height. This paper presents a novel inspection system based on the application of the stereo technique. In our single- camera setup, stereo views are obtained by rotating the IC chip under the camera. Wire bond edges are detected using Canny's method. These are then linked to form curves, and matching is conducted between two curve lists of a stereo pair. Results of local matching are combined to find the global optimum. The depth measurement of our system has an accuracy of 14 micrometers , which corresponds to one-fortieth of the total range of wire height.

In this paper, a new model-based tracking algorithm is proposed for tracking rigid objects in 6 degrees of freedom. Only one calibrated camera is used in the approach which can handle the motion of objects with known geometry. Information in 2D images from the camera would conduct motion in 3D space. The useful image features are contour edges of object to be tracked. The matching process includes two aspects of: (1) feature extraction using local minimum energy and (2) global matching of known 3D models against the projected features. The algorithm is robust to change in lighting and background. The small motion hypothesis is used for fitting feature energy which is defined as the negative absolute value of the edge strength. An autoregressive AR(1) model is employed for detecting incorrect matches in terms of the feature energy. We have found a new invariance-based method to eliminate false matches caused by strong shadow or occlusion. The invariance is the ratio of trigonometric functions of the angles formed by a polygon. Both performance analysis and real object tracking show that the proposed algorithm is effective and robust.

In modern semiconductor and optics industries, there is a strong demand for a highly sensitive and non-contact surface profilometer. This paper describes an optical heterodyne surface profiling interferometer with automatic focusing which has been developed recently. The essential feature of the profilometer is a newly designed common-path method to minimize the effects of environmental conditions. A powerful signal processing scheme is also developed, which includes three parts: automatic voltage control, phase measurement and automatic focusing control. All these make the repeatability and stability of the profiling interferometer greatly improved. The height resolution is 3.5 nm and lateral resolution is 0.4 micrometers . Experiments show that the profiling interferometer is suitable for on-line use.

One of the areas in which a great deal of research is taking place is the application of image processing techniques for automatically detecting moving objects. The common object detection techniques used by several researchers are based on background differencing. However, in practice, the accuracy of this technique mainly depends on the methods used for background updating and the thresholding techniques. Edge-based image detection is generally more effective than background differencing and has been used by few researchers for detecting moving objects. The conventional gradient-based edge detection operations have found wide acceptance in image processing applications. However, morphological edge detectors have shown better performance than conventional edge detectors while having a lower computational cost. This paper describes a novel method for object detection based on morphological edge detection techniques. This method eliminates the background updating and has been implemented in real-time on a low-cost Pentium-based microcomputer system.

Periodic corrugations in the form of gratings are an important component in integrated optical devices. In most cases, the period of the corrugations required is of the order of a micron or less and hence conventional photolithography could not be applied. In this paper, we discuss the development of a holographic interference exposure system based on a He-Cd laser of wavelength 325 nm. The laser light is kept as one beam as it transverses through the various optics. Upon incidence on the exposure assembly, the wavefront reflected from a mirror and that incident directly onto the photoresist-coated sample will interfere to generate the required interference pattern on the sample. Gratings with periods of 320 nm to 1.69 micrometers have been fabricated on positive photoresist-coated silicon samples. The periods were analyzed by both Atomic Force Microscopy and optical diffraction measurements. The periodic patterns have also been transferred onto glass substrates via etching. The measured grating periods are in good agreement with the designed values.

A phase-shifting Twyman-Green interferometer has been constructed. Using three consecutively captured interferograms, the phase profile of a reflective surface can be determined. Results using various fringe processing techniques are compared. These methods include uniform averaging, Gaussian mask and spin filtering. For simulated fringes superimposed with random noise and fixed-pattern noise, it has been observed that a combination of weighted averaging and spin filtering could generate the best results. The computerized system has been applied to the measurement of the form errors of a silicon wafer and a cosmetic mirror, respectively. The root-mean-square error of the wafer is determined to be 11.13 nm.

A novel solution for image pattern identification based on the description of linear structure is proposed. The concept is currently applied to the classification of documents such as business forms and logos. The geometric layout of objects such as lines, text and spacing is converted to a 1D string representation. A novel, generic and quantized string format is proposed for the system. Very encouraging results have been obtained and the technique can be used for a wide range of applications and extended to handle patterns including electronic circuit diagrams without obvious linear features. The use of strings facilities quick and robust measures of similarity between two patterns and a quantifiable tolerance of segmentation inconsistencies is possible. In addition, no training is required.

A novel solution to the problem of correcting leaded semiconductor packages damaged during test and handling is presented. The elimination of the high costs of reworking and maintaining lot integrity is accomplished by integrating a highly programmable lead conditioning system with the traditional lead inspection system. The equipment accommodates gull and J lead forms via a quick change lead conditioning tool. Devices ranging from 10 mm to 40 mm, square, rectangular, two or four sided can be effectively conditioned with no increase in setup time or decrease in overall throughput. The effectiveness or device recovery rate of such a system is documented with data from semiconductor manufacturing and printed board assembly environments. While previous methods relied on huge amounts of hand labor and very expensive hard tooling, the programmable conditioner allows a high product mix within the setup and changeover time of the inspection equipment. Data demonstrating the effectiveness and versatility of a `one-lead-at-a-time' conditioner is presented. Although the throughput is less than for inspection alone, the total cycle time and throughput are actually higher than for process configurations which separate inspection and conditioning. Conventional processing flows impose severe cycle time delays, hand labor costs and re-inspections that are eliminated in the integrated system. The major factors of determining the payback and justification for an integrated system are also discussed. The economic factors discussed include direct labor, capital cost, tooling cost and cycle time. Systems justified to date have done so on the basis of a single factor, the value of the recovered units. Intangible factors are also identified and discussed.

Visual inspection has long been a necessary method of quality control in Printed Circuit Board Assemblies (PCBA) manufacturing. The characteristics of electronic assemblies have changed substantially over the last decade. Todays high lead count, fine pitch SMT components are becoming even more difficult for humans to inspect at the same time automated inspection systems have become reliable than manual inspection and are now accepted as valuable tools for producing high quality PCBA products. The basic requirements of an automated inspection system remain same in all PCBA manufacturing but the type of the automated system (off- line/on-line), where applied in the production flow, entire boards or only on a sample basis, inspection coverage (100% or partial) vary between different PCBA manufacturers. In PCBA manufacturing the emphasis is more in the electrical functionality of the PCBA than in it's appearance. It is nearly impossible to impose stringent specifications in the appearance of the components and other materials used in PCBA manufacturing. Due to the large number of component/PCB supplier and wide variations in materials and processes the challenge in successfully automating the inspection process is the variability in the appearance of components on PCBA. But in a high volume PCBA manufacturing where fewer board types are running in large volumes for long periods of time, the variability in component appearance can be controlled much better than a low volume PCBA manufacturing where more types are running in low volumes for short period of time. This paper discusses the development and implementation of a low cost flexible automated inspection system for PCBAs. The system can detect over ninety percent of visual defects on PCBAs. The key features of the system are quick and easy set-up, capability to inspect different types of board and quick change over between different boards and low cost.

This paper presents an investigation into the color space and the recognition methodology for developing a non-contact approach to reading the value of a resistor by identifying the color bars from its color image. In this work, several color spaces for measuring colors are evaluated and compared. The uniformed color space known as the CIE-LAB color space is found to exhibit good properties for color recognition task. Hue(H)-Saturation(C)-Intensity(V) color space, which are calculated from CIE-LAB color space, provides an intuitive measure of color. Two color identification methods are investigated. The first method uses the Bayes classifier which defines color boundaries based on the variances of colors. The second method uses predefined fixed color region boundaries. A color is identified if it falls within a color region. A combination of both methods can achieve reliable recognition of color bars and hence correct resistor value can be derived.

Based on optical non-contact sensing and machine vision technology, an optical non-contact measurement instrument for measuring blind holes, as well as through holes, has been developed. By projecting a precision optical target on the internal wall of hole and sensing its reflected image, the instrument can be used to measure the hole's diameter, ovality and cylindricity. An optical selective amplification system was developed that greatly increases the sensitivity of CCD camera. Experimental test shows that the sensitivity of the laboratory apparatus is about 3 micrometers /pixel and the measurement accuracy is about 10 micrometers .