Defected product detection in industrial assembly lines using image processing technology with limited data

From past to present, with the beginning of the industrial revolution, mass production lines have been developed differently for each sector. With the introduction of mass production, it is not possible for all products produced on these lines to be defect-free due to mechanical and technical errors. For this reason, different solutions have been tried to be produced within the quality control processes. Traditionally, defect detection on production lines is done by human observers. This method is a tiring job that requires high attention and is a process prone to errors. Additionally, detecting errors using this method can be time-consuming and costly. With the rise of artificial intelligence (AI) technologies, new solutions have begun to be developed for error detection. Artificial intelligence-based systems stand out with fast and accurate error detection. However, current AI systems require large amounts of data to work effectively. Moreover, the performance of these systems decreases, especially in production environments where little data is available or constantly changing. As a result, anomaly detection with little data has become a separate field of study. In this study, a data set was created with 400 error-free metal beverage cans and 10 anomalies. As a method, a hybrid model combining the YOLOv8 segmentation model and RegAD along with the RD4AD and RegAD models was used. No pre-processing was performed on the images in the data set. The hybrid model proposed in this study, the YOLOv8 segmentation model, was trained on images containing anomalies for 50 periods and the results were measured on the data set using the 8-shot method for RegAD. As a result, it was observed that the hybrid model proposed in this study showed more than 10% more semantic success than other methods on our customized data set.