Neural Network Approach To

This paper discusses the use of an artificial neural network in an automated IC package wirebond inspection defect classification system. A neural network was applied to the inspection of wirebonds in high pin count (>300 pin) IC packages. This study required sample acquisition, image capture, image segmentation, data compression, feature extraction, and training a neural network. Simple image processing algorithms and semi-automated image processing tools were used to analyze the final database of 1404 pad images. The categories with more than 100 samples had a >90% classification accuracy, other categories had 75 to 78% accuracy. Commercially available image processing tools are necessary to make this process manufacturable. BACKGROUND AND MOTIVATION This paper discusses the development of a neural network solution for automated IC package wirebond inspection, as a feasible alternative to current visual inspection. Visual inspection of any form is time consuming expensive, and subject to interpretation and fatigue. As the objects being inspected become smaller, denser, and more complex, the inspection process becomes more prone to human error. Traditional machine inspection has been expensive, time consuming, and not as effective as human operator inspection. The result has been continued human inspection or no inspection. Wirebond inspection for high pin count (339 to 600 pins per package), high density devices (>0.75 Million transistors per IC) can be used to rework devices improving yield and profit. Inspection error lowers electrical test yield and can cause non-testable field returns. Automated inspection in industry has increased in the past 5-10 years, automated areas include: autobody metal and paint, clutch drivers, and PC board solder joint inspection. Automation development is taking place in PC board solder joints, IC wafers, chili-pepper harvesting, IC wirebond inspection, turnkey wirebond and die-attach systems, and 3-D imaging for accept/reject processes [1]. Improvements in lighting systems, microprocessors, image processing systems, and camera technologies have opened the field to replace manual inspection with automated inspection [2]. Specialized lighting systems are required to obtain clean predictable images. Development of high speed, low cost, PC compatible image processing hardware has expanded the automated inspection field [2-3]. Systems incorporate fast specialized image processing hardware to perform the image capture, rotation, region of interest selection, image enhancements, feature extraction, measurements, 3-D mapping, and data compression processes; thus accomplishing precision vision analysis. Software and hardware neural network tools are now available with the high speed image processing boards. In a comparison with traditional automated vision systems, Huang [2] found that the neural network took longer to train than area calculations, boundary following, or histogram calculations. However, the accuracy of the neural network and the orientation flexibility made it a more effective classification tool.