Image-tile subtraction shortens ceramic-tape inspection time
In the production of ceramic component carriers for the electronics industry, conductive metallic material is printed onto green ceramic tape. After printing, the conductive material is fired and then plated with gold. Because voids in printing and excess printing materials can cause open conductors and shorts in the final product, determining whether voids or excess materials are present before firing makes this production process more cost-effective.
To accomplish this, researchers at the Coors Electronic Package Co. (Chattanooga, TN) and Oak Ridge National Laboratory (Oak Ridge, TN) have developed a VME-based system that can detect flaws as small as two-thousandths of an inch on parts up to 8 ¥ 8 in. To accomplish the 2-mil/pixel resolution, a 2k ¥ 2k MegaPlus camera from Eastman Kodak (San Diego, CA) is mounted on an x-y axis stage from Lintech (Monrovia, CA). The camera is moved across the ceramic component carriers and acquires four 2k ¥ 2k images, which are then assembled into a composite 4k ¥ 4k ¥ 8-bit image.
"The use of a golden-template reference-based inspection was an obvious choice for this inspection," says D. Lamond of Coors. In this approach, images of printed parts are precisely aligned with a flawless template and are then compared by pixel subtraction. Differences between the images appear as nonzero pixels.
To produce template images in BMP format, CAD files for the component carrier artwork are converted into BMP format using an ASM 501-DXF-to-Gerber translator from Artwork Conversion Software (Santa Cruz, CA). Once produced for a given part, the data are saved for later re-use by the image-processing system.
To reassemble the 4k ¥ 4k image, the individual 2k ¥ 2k quadrant camera images are first captured and then registered with fiducial tie-points to the original CAD image using a miniwarper module attached to the VME-based MV200 pipelined image processor from Datacube (Danvers, MA). Using second-order bi-quadratic mapping, the 2k ¥ 2k image is warped using nine image "tiles" of 670 ¥ 670 pixels each, resulting in a composite 4k ¥ 4k image of 36 individual tiles.
After subtracting the captured image from the template images, the difference is then multiplied by the mask complement image to reduce the number of pixel differences. Defects are then binarized using a thresholding algorithm, and the centroids of the defects are located and evaluated using a defect specification map.
"Although overall system throughput varies between 9 and 20s," says Lamond, "the time required for template subtraction is 0.85 ms. In this system, the rate-limiting factor is the image processing performed by the host CPU, a VME 167 from Motorola (Austin, TX) running the Lynx (San Jose, CA) operating system. However, with the emergence of newer and faster MV200 host processors with greater RAM, the overall inspection time should be reduced," he says.
