OCR for Unreadable Damaged Characters on PCBs Using Principal Component Analysis and Bayesian Discriminant Functions

In the last few decades, new computer vision technologies and image processing techniques have been very important in the improvement and automation of manual processes in many technical areas, e.g., in the semiconductor industry. In this paper, we propose to change the actual pattern matching methods implemented to have optical character recognition by the use of principal component analysis method to extract the principal characteristics and features of damaged or unreadable numerical digit characters from images on printed board circuits (PCBs) and compute linear and quadratic Bayesian discriminant functions to classify and find the correct numerical character that corresponds to those features. In the first step of this work, grayscale color images are acquired from a charge-coupled device (CCD) camera, then image segmentation is manually computed to create a dataset of 500 matrix images for the character digits from 0 to 9. Then, a feature extraction method is applied to get the principal components that will be used in the character recognition state. Finally, our results show that applying Bayesian linear and quadratic discriminants to the principal component features can improve optical character recognition (OCR) detectability of damaged characters from actual 95 -- 97% to 99.88% in early tests. This suggests to us that the problem probably follows a linear model where linear hyperplanes separate decision regions with satisfactory (almost no) errors.