“Production lines are not designed to produce defects.”

The basic method of defect detection using deep learning is anomaly detection, where the system purely classifies parts as OK or NOK, based on their deviation from the few images of an ideal part.
Apart from this method, other techniques are used in visual inspection (VI) to identify, localize and classify defects, where a substantial number of images representing such defects and features are required for training purposes. More soever, precise and consistent labels are key to ensuring a high performing deep learning model.
To create a balanced data set for training and validation, one needs both good images (OK) and bad images (NOK). Since ideally no components with defects are produced in manufacturing, one often finds a shortage of defect images (NOK) available for training and validation (within a reasonable time period).