Overview
To prevent the sales of defective products, every single part must be inspected. In order to automate the time-consuming and costly procedure of human inspection, an image is taken of every part and subjected to algorithmic analysis. Instead of designing individual algorithms for specific defects, a statistical classifier can be trained. This requires a set of images of defective and intact parts, along with labels. The labeling effort is reduced by requiring only "weak" labels that a human expert can specify with little efforts. More specifically, the definition of "weak" labels used here is as follows: the human expert is instructed to provide labels not at pixel-precision, but by generously outlining a defect in the image with an ellipse. The labeling effort can be reduced even more by so-called active learning approaches, which aim at obviating the labeling of parts with little novelty for the classifier. Current research also deals with composite parts, typically with complicated geometry, by incorporating structural information into the learning scheme.
Methods
The requirement of generic applicability of this system to many different defect detection problems requires the use of a combination of methods from image processing and machine learning. The image processing part involves the extraction of features encoding the appearance characteristics of defective and non-defective image regions, while classification algorithms from machine learning are trained to make predictions based on the extracted features of previously unseen images.
Classifier Output
The result of the effort is a defect probability map from which a manufactured part can be classified as defective or not.
