Quality control represents a crucial step in any industrial production cycle. The motivation can be diverse such as integrity of building goods, safety of foods or flawless cloths. However, the challenges a manufacturer faces are usually very similar. Ideally these control steps can be carried out quickly, are cheap and highly reproducible and have a vanishing error rate. Solving all these demands simultaneously can be an extremely hard task.
Many of these control tasks rely on visual information like an image stream from a conveyor belt. Evaluating product photos in a standardized manner can be problematic due to data variances, e.g. visibility of different product parts, variety in illumination, image quality. Here, rule based approaches using classical computer vision algorithms can be helpful but rapidly reach their capability limits. Furthermore, special evaluation processes might need a manual assessment. These are especially time consuming and often need employees with unique expertise. In addition, analysis steps which are carried out by humans tend to be less standardized and such results can fluctuate drastically.
In this context deep learning approaches can be extremely beneficial either as a stand-alone solution or in a hybrid system, where they act as an assistant. Customers can pick from a large variety of different algorithms to solve frequent tasks like anomaly detection, product categorization, defect detection or part-geometry analysis. One of the big advantages of machine learning methods is the improved generalizability. This means that the capability to capture image variety and product outliers is significantly improved and thus surpasses the performance of rule-based approaches. Some assessments, like relating geometric structures, become tangible only in this way and are in reach to being automated. Additionally, control processes can be further standardized to increase the comparability between factories, employees or regions.