Method for identifying target anomalies in a digital image

Abstract

A novel approach for identifying target anomalies (10) in a digital image using convolutional neural networks (CNNs) is described. Typical CNN involves a classification process that involves large training data which may still result in false negatives and false positives. To address the matter, the current invention includes a step for delineating contours of feature descriptors representing non-target anomalies. These feature descriptors are subsequently filtered out, allowing the CNN to focus on the target anomalies (10). The anomaly tolerance of the CNN can be statistically quantified to achieve the desired outcome. By reducing the incidence of false negatives and positives, this method shows promise for routine applications such as identifying defects in printed circuit board tests, particularly for detecting faulty holes. The method uses a relatively smaller dataset compared to the dataset typically required for CNN.