Abstract
Automated visual inspection algorithms for printed circuit boards have long been focused on geometric defects of conductive paths that comprise regularly straight line and circular arc segments. In this study, we aim at defect inspection of substrate bond pads consisting of rotated and deformed shapes. A non-referential, self-comparison machine vision scheme is proposed in this paper for shape defect detection of bond pads. It is invariant to translation and orientation, and is tolerable to shape deformation of faultless objects. A shape reconstruction technique based on Fourier transforms is first applied to restore an inspection object with varying smoothness. Two discrimination features are then extracted from the point-to-point distances between the original and reconstructed shapes. The first discrimination feature measures the global irregularity of the whole boundary of the inspection object, and the second feature measures the maximum local deviation of the object boundary. Experimental results of numerous real bond pad samples have shown that the proposed self-comparison scheme with the two discrimination features can well separate the clusters of faultless and defective samples.