Inspection detects the slightest defects Friedrichs & Rath GmbH (F&R) is a specialist for precision parts made of thermoplastic materials and manufactures a high proportion of these parts for the automotive industry. In order to meet the high quality demands, especially of customers in the automotive industry, while at the same time producing large quantities, F&R would like to introduce a learning, precise automated optical quality control system. The inspection system should be able to reliably detect even the slightest defects and, if possible, classify and categorize them. The goal is an inspection application that can be adapted to different types of plastic parts with little effort, even by non-specialist employees.

In this first evaluation study of learning image processing methods, Fraunhofer IOSB-INA first recorded the requirements of F&R with respect to defect types and representative inspection sample types. In contrast to the classic parameter-driven image processing learning methods are based on sample images and corresponding classifications (eg "OK", "NOK").



It therefore does not require expert knowledge in the field of image processing. For this reason, test samples of six different types of plastic parts were provided for the evaluation. These were classified in advance by F&R either as "OK" or with a defect or as "NOK". Fraunhofer IOSB-INA first photographed these test samples using a generic camera and exposure setup. Since each plastic part type has its own shape and potential defects occur at part type-specific locations, an individual frame was created for each part type using additive manufacturing . The images created were used to test the approaches of classification as well as anomaly detection learning methods. Classification attempts to recognize learned patterns. Anomaly detection tries to identify deviations to a learned reference image. For the evaluation, 70% of the recordings were basically used for training, 15% of the recordings for automated validation / verification and 15 additional percent for the final test.

.

In classification, OK as well as NOK images are considered for training and validation. For anomaly detection on the other hand, only OK images are used. The test is also performed with OK and NOK images in the case of classification. In the case of anomaly detection, testing was done with all NOK images and 15% of OK images. The results of the test were very promising. Detection rates of 96-100% were found using anomaly detection and 100% for all part types using classification. These high detection rates required a minimum number of 30 test samples per part type and a systematic classification of the images. Based on the good results, F&R plans to design and build a generic automatic testing machine that includes the evaluated method. This testing machine should be able to test different plastic parts and thus enable 100% testing for production lines, which was previously not economically feasible. 

.