Open AccessSupervised learning vs. unsupervised learning: A comparison for optical inspection applications in quality control
Author(s) -
Jan Lehr,
Jan Philipps,
Van-Thanh Hoang,
D von Wrangel,
Julia Krüger
Publication year - 2021
Publication title -
iop conference series. materials science and engineering
Language(s) - English
Resource type - Journals
eISSN - 1757-899X
pISSN - 1757-8981
DOI - 10.1088/1757-899x/1140/1/012049
Subject(s) - computer science , unsupervised learning , artificial intelligence , machine learning , anomaly detection , quality (philosophy) , control (management) , novelty , supervised learning , semi supervised learning , field (mathematics) , pattern recognition (psychology) , artificial neural network , mathematics , philosophy , theology , epistemology , pure mathematics
For the establishment of a successful quality management system in companies, the quality control of e.g. newly produced goods or the return of old and used parts is an essential component. One solution for this is the optical inspection of the surface of objects with the help of image processing algorithms. Using the case study of printer cartridges, this paper evaluates the extent to which different methods of machine learning can contribute to a successful quality control. Established methods of supervised learning have the advantage that they are already proven in many applications and have a very high detection accuracy. However, they require a lot of labelled training data and this high effort also means high integration costs. A new approach is a data-reduced variant from unsupervised learning. Here, the algorithm is trained only with defect free objects, for example as they come to a large extent from the production. If the objects are defective, the method from the field of anomaly detection or even novelty detection detects something that is different from the learned norm. This has the advantage that not all defects have to be known beforehand. And this in turn avoids acquiring a large amount of training data for each of these defects. This paper compares the effort required to acquire training data and compares it with the detection accuracy of the different methods in order to give an assessment of the extent to which the use of unsupervised learning methods is beneficial. Newly produced and used printer cartridges are used for this purpose. Image data is acquired from 18 different printer cartridge models. Afterwards they are fully annotated (labelled). A smart separation into training, validation and test data allows the training of supervised and unsupervised methods as well as a complete evaluation regarding the effort for data acquisition, annotation and detection accuracy of the defects. Finally, an outlook for chances and risks of the respective procedures is given.