|
The increased quality demands on our electronic equipment has resulted in more and more production lines being equipped with AOI (Automatic Optical Inspection) equipment, to detect defects, provide feedback to help the user eliminate the cause of those defects and ultimately control the line toeliminate the defects. The increasing use of smaller components, 0201 and even 01005 demand greater and greater resolution from the AOI. The need to reduce costs means that line speeds are increasing and the AOI needs to be faster in operation so as not to represent a bottleneck in the line. These two demands are in conflict and generally create a dilemma for the AOI designer. Alternative approaches
Several alternative approaches are offered to solve this dilemma with varying degrees of success. A classic vision system is composed of an image capture device, (in high end AOI this is a high resolution CCD image sensor), a lighting system, and software vision algorithms that are used to interpret features in the image and thus identify defects in the assembly. The resolution of any vision system is optically determined by the camera sensor’s physical pixel size and the objective lens magnification. This gives the pixel size parameter of the system, which is in fact the size that 1 camera pixel represents in the scene in the real world. The presence or absence of any detail in the image that is smaller than this pixel size cannot be detected. We often see the advances in digital camera technology with 10 mega pixel being the current state of the art. Rugged and reliable commercially available image sensors suitable for continuous industrial usage however are typically in the 4 mega pixel range. A 4 mega pixel unit will have 2048 x 2048 pixels in a detector 15.15 x 15.15 mms which gives a camera pixel size of 7.4 micron. So if our inspection field of view were 15.15 x 15.15 mms, (A 1:1 objective lens) then our image pixel size would be 7.4 micron. And if our inspection field of view were 45 x 45, then our image pixel size would be approximately 22 micron. As can be seen, for the same detector, increasing the field of view decreases the detection pixel size. The first approach then would be to merely rely on the optics, increasing the objective lens magnification of the system. In some instances, this is an acceptable option. However reducing the field of view means that more images need to be captured and this increases the cycle time. Ideally we need to have a large field of view and high accuracy to analyze a full range of components from large QFPs down to small 01005 chips. Another approach is the use of multiple cameras to cover the same area with an increased resolution. Ensuring and maintaining common calibration between the cameras becomes a concern. In most cases this leads to higher equipment cost, higher cost of ownership and therefore does not constitute the most efficient response to industry requirements, even though accuracy reached could open the window of the assembly process control. Yet a third approach uses a switchable objective lens to increase magnification and provide greater resolution over critical areas, again this solution increases inspection time and also relies on mechanical switching systems which could degrade system repeatability. The best possible outcome to serve the demanding industrial requirements is the combination of leading edge optical technology utilizing the best commercially reliable camera detectors, high quality optics and subpixel capabilities of advanced image processing algorithms. Subpixel techniques enable the location of object features with a higher resolution than the pixel grid. Additionally to assist with improving the process and achieving optimum fist pass yield, the AOI systems are required to measure objects rather than just detecting presence or absence. Thus the ability to locate objects or transitions very precisely is critical. In this way it is possible to control the process rather than just monitoring defects. Subpixel techniques provide precise measurements using continuous scales instead of discrete grids aligned on and limited by pixels. Locating vectors compared to counting pixels
As seen in figure 2, pixel based technology will try to recognize an object by counting pixels and comparing it to a stored image. This method is relatively slow and is adversely affected by changes in color, background, sizeand rotation. Vectoral imaging approaches the problem with a completely different method in order to overcome the problems of grid based pattern analysis. It converts the pixel grids provided by the image sensor into geometric features. Vectoral imaging extracts and locates features (object contours), converts them from the pixel grid to vectors, and analyzes them as geometrical shapes. A previously stored set of vectors (either extracted from an image or generated from CAD data) is compared to the vectors extracted from the run time image. The tool has an accuracy of up to 1/40th pixel in position and 1/50th of a degree in rotation on defined targets. On the smallest targets such as 01005 components, that represents only 5 x 10 pixels in size (with a 26um system pixel size), but the accuracy is close to 1/20th pixel (about 1 micron). The vector extraction uses a multi-directional subpixel edge location algorithm (similar to the one described in figure 1) giving a set of vectors that represent the object boundaries. Each vector has realvalued coordinates and orientation with subpixel precision. After the vector extraction, all subsequent matching steps operate in the mathematical domain, to correlate a set of trained vectors with another set of run-time vectors. The set of vectors that defines the object can be easily and efficiently translated, rotated, scaled, stretched or distorted with no loss in fidelity. Accuracy is unaffected by the orientation, size and even shading variation. This gives very fine location accuracy, as the template vectors can fit the run time vectors optimally. Each individual vector contributes to refining the position result. The results are not dependant on the pixel grid anymore, thusgiving real-valued measurements. Vectoral imaging is widely used for accurate component inspection and ViTechnology also employs a wide range of other vision processing tools. Some of these tools are based on advanced edge locating tools, to detect component body and leads. Beyond image processing algorithms that locate features, another key to obtain good subpixel accuracy is the calibration algorithm. The image that is processed by the feature detection algorithms was obtained through a complete optical system, including an objective lens. This gives a flat representation of the three dimensional scene, and introduces optical distortions and errors. Thus the conversion from the pixel grid to real world units can not be a simple linear equation. Advanced non-linear calibration algorithms make the most of subpixel locating tools. Combined with a custom designed telecentric lens ensures a completely orthogonal image over the whole field of view. Conclusion
AOI inspection in SMD electronic assembly lines provides an enduring challenge for AOI vision applications. Superior subpixel technology enables the location of component defects with both the highest accuracy and highest speed available. The combination of state of the art image detectors and custom designed optics, together with programmable lighting and powerful subpixel technology offers the best solution available for today’s and tomorrow’s SMD assembly lines around the world. Figure 2 Figure 3 Figure 4 Vectoral imaging extracts component boundaries and analyses them in a mathematical domain to locate the component in the picture with a precision higher than the pixel grid. | Subpixel world: From simple tools to advanced pattern matching Several image processing tools have been developed which can locate features or objects, and measure distances. To illustrate the concept in one dimension we will use the ‘edge tool’. • Edge detection
Basic edge detection tools detect features along one axis, which is often aligned with the pixel grid. They use one row of pixels, and interpolate the grey levels to enhance the resolution looking for a transition of light to dark. A more sophisticated approach is to apply a derivative operator at many points across the pixel row at 90 deg to detect the transition point. The real and more accurate position of the transition is located right at the contrast slope inversion, when the derivative reaches its maximum value. Using this technique at many points along the pixel row, produces a more accurate transition edge than can be achieved with simple grey scale correlation, and not now limited by the pixel size. Resolution of ± 0.1 pixel or better can be achieved. • Pattern matching
Pattern matching tools, also known as normalized correlation tools, work with a portion of image (pattern or template) that is previously stored in memory. The tool compares like-sized subsets of the run time image with the template, in multiple positions, and computes a correlation factor. This gives a probability of the pattern presence for each location in the runtime image, with an accuracy of 1 pixel. As the template image typically contains much more pixels than an edge detector, re-sampling the image to enhance the resolution would just dramatically increase the required memory and search time. Most of the pattern matching tools propose a fine search option that hardly gives a quarter of pixel resolution at the cost of requiring additional memory and search time. Figure 1 | About the authors: Cathy Combet is the Technical Product Manager, and Régis Braisaz is in AOI Research &Development at ViTechnology, Saint Egrève, France. | 
