By Gary Wagner Imaging Technology Inc. Bedford, Massachusetts
Overview As shown in Figure 1, the typical machine vision system uses one or more video cameras to precisely position or inspect parts. Proper lighting of the scene is nec-essary for the camera to reliably see subtle defects or differences in contrast. A lens focuses an image from the part being inspected onto the camera's sensor. The signal from the video sensor is digi-tizedÑ converted from a continuous (analog) signal into samples called "pixels." These pixels represent light intensity (or some other measure) at points on the part. An array of pixels is called an "image." The image is typically manipulated by general purpose processors, such as Pentium-class personal computers. Two types of manipulation are involved: image processing and image analysis. First, image processing converts one image into another image. Processing operations are used to select or amplify important features in an image, such as the edge of a package. A feature, roughly speaking, is a component or attribute of an object such as its color, intensity, texture, or edges. Second, image analysis converts an image into measurements, such as the package's size and location. For example, image analysis is used to characterize the size and shape of particles trapped on a PWB to determine whether a high-quality solder joint can be produced. In machine vision, measurements such as color, size and distance between edges are used to evaluate various package parts and reject those with defects.
System Components Machine vision systems acquire and process images and use the collected information to report on (or control) a process. The primary components involved include cameras, lighting, one or more vision controllers or frame grabbers and machine vision software running on a computer. Lighting Four primary types of lighting can be used in machine vision applications: front lighting, backlighting, structured lighting and strobe lighting. The goal is to find the combination of techniques that results in the best possible image. Front lighting illuminates the area in front of a part. The result is a pattern that is created by the part intersecting the light. Backlighting places the light source behind the object of interest so that a silhouette is created. Structured lighting uses sheets of light to determine the shape or dimensions of a part by observing the pattern made when the part intersects the structured light. Finally, strobe lighting allows high-speed processes to be monitored by providing a short burst of light, typically in the microsecond range. Within these four categories are a dozen or more specific types of lighting techniques. For example, diffused lighting is employed when image contrast is high and edges are distinct, such as when traces contrast with the background wafer. A special way to implement diffused lighting is with a light tent. This technique helps eliminate glare, such as when imag-ing ball grid arrays, by bouncing light off a surface placed in front of a part. Cameras Today, most machine vision applications use analog cameras to capture the image of the part being inspected. While digital output cameras are available and may produce better signal-to-noise ratios, the benefits are not always cost justified over lower cost analog cameras. Ultimately, the application's inspection rate, algorithm details, and physical constraints will determine the speed, res-olution, and type of camera used. While RS170 or CCIR cameras are common, the need for more detail and faster acquisition rates is rapidly driving machine vision in the direction of higher resolution, pro-gressive scan cameras. Frame Grabbers A frame grabber is an image capture board that connects to either a proprietary processor or off-the-shelf computer (Figure 2). The board takes a set ("frame") of image data from a camera, converts it from analog to digital, and stores it in memory for further processing. A primary difference between com-peting frame grabbers is their memory architecture. There are two basic kinds: The FIFO-based and the on-board memory-based. A FIFO-based design streams data into a 1K or 2K temporary buffer at the output rate of the camera. Data is sent from the FIFO to the host computer's memory after the FIFO fills. A problem with this type involves data-rate mismatch between the incoming data (usually 0 Hz to 50 MHz, slow scan video to very fast scan video), and the computer's bus architecture, which, in the case of the PCI bus, must transfer data across the bus at 132 Mb/second. On the other hand, memory on the frame grabber, coupled with a DMA (direct memory access) engine, allows data to simultaneously enter from a camera sensor and be transferred to the host computer's memory with very little over-head from the CPU. The difference in CPU loading can be as dramatic as 1% from a frame grabber with on-board memory vs. 42% from a FIFO frame grabber. With 40+% CPU loading, there is little time left for other computer processes, like image analysis. Vision Software When considering machine vision software, first assess how much of the software inte-gration task you are willing to assume. The choice of whether to write a custom algorithm or use an off-the-shelf package depends on the uniqueness of the application and the time constraints of the project.
In general, using software packages written specifically for machine vision not only speeds the completion of the project, it usually results in a faster, more reliable solution. Common Bottleneck One of the most common bottlenecks is the integration of the host computer-based software with the frame grabber. Because of this potential problem, strong consideration should be given to the maturity of the interface between the frame grabber's software and that of the image processing package being considered. Basic frame grabber software function-ality should, at a minimum, include methods for triggering, strobing, frame reset, data transfer to host, image display, and a way to dynamically tweak the camera interface. These are invaluable features when first setting up the system to acquire an image. Once the desired image is acquired and delivered to the host memory, the task of actually solving the machine vision problem begins. Here, again, access to a powerful image processing library designed for machine vision (vs. medical or general purpose image analysis), can greatly improve performance, decrease develop-ment costs and improve success rates. Solving Problems
Machine vision has been particularly successful in the evaluation and testing of electronic components. These devices are precisely built and with few variations, so a machine vision system can easily inspect for defects based on a defined set of limits for a specified set of measurements. Figure 3 shows a machine vision system measuring the spacing of leads on an IC package. The leads are illuminated so that they are bright white against a black background. A threshold operation on the input gray-scale image creates a binary (or bi-level) image where the leads are all one intensity value, say 255, and the back-ground another value, say 0. Image analy-sis is used to find each lead and measure its area and approximate position. Using this information, precise meas-urements of lead position are made on the gray-scale image. These measurements can typically be made from 1/4 to 1/16 of a pixel (or more) with special conditions. If the field of view is 50 mm and an image with 640 columns by 480 rows of pixels is used, then with 1/10 pixel accuracy the measurement limit is about 0.01 mm. Evaluating Bond Pads Figure 4 shows evaluation of the bond pads on a small PC board. (The pads are used to electrically connect an IC to the package pins.) The gold bond pads are bright white. As before, the input gray-scale image is brought to its threshold to generate a binary or two-level image; then a series of image processing operations, known as morphological operations, is used to gauge pad size while looking for any breaks. The photos in Figure 5 show part of the process for a small area near the top of the previous image. The left image is a magnified view of the original image; the next is the result of bringing the original to its threshold. A morphological closing is used to "fill-in" any cracks in the threshold image, resulting in the closing image. Subtracting the threshold image from the closing image yields the result image, which clearly shows the gaps or open circuits in the bond pads. This part is defective! New Technologies The above examples, along with most machine vision applications, make heavy use of a traditional mathematical tool for finding and aligning patterns: normalized grayscale correlation. This correlation has been used for years within the machine vision industry as the main vision align-ment tool. In the past, it has met the demands of most alignment tasks, but today there is a big push for machine vision software that can handle a much wider variety of process variations. While normalized grayscale correlation is an accurate pattern-matching algorithm under ideal conditions, its accuracy degrades with variations in contrast, rotation, scale and partially degraded and occluded patterns. For those reasons, the need to provide the industry with fast, flexible and accurate pattern locating tools has become a must. For example, wafer manufacturing presents a difficult challenge for machine vision. In pattern wafer alignment, many changes are made to the material as it proceeds through the production process. To accommodate these changes, machine vision tools are required to perform under a variety of challenging conditions, including:
Under these conditions, the vision tool must locate the position and orientation of the wafer within 1 µm in both the X and Y directions, and within 0.1 degree theta of rotation. To accomplish this positioning, new, adaptive pattern location tools have been developed using geometric search tech-niques. These new tools are able to over-come the kinds of variations that affect the appearance of a pattern. Once the desired pattern is located to within a pixel, a sub-pixel calculation is performed using neural-net-based techniques to compute high sub-pixel accuracy. The result is a robust and precise pattern recognition tool that allows man-ufacturers to handle the variations inevitable in most packaging processes more easily (Figure 6).
Conclusion Machine vision has finally come of age as a mainstream tool for performing precise, repetitive and high-speed package inspec-tion. Furthermore, this technology promises continued advancements that will enable new and greater levels of manufacturing quality, flexibility and cost control. Mr. Wagner is President of Imaging Technology Inc. [gwagner@imaging.com] |
