Applying Machine Learning Techniques for Anomaly Detection in Wooden Plank Images

Abstract: Anomaly detection is an important first step of quality control in manufacturing processes. In wooden planks, anomalies such as broken knots and resin pockets can lower the quality of the final product. With the help of machine vision, inspections can be made faster, at higher accuracy, and at a lower cost. Therefore, this Master's Thesis project aims to explore different machine vision-based machine learning methods for anomaly detection in images of wooden planks. Both unsupervised and supervised methods were used. The evaluated unsupervised methods were two variations of student-teacher frameworks, while the supervised methods were different semantic segmentation models. The evaluation results showed that the pre-trained DeepLabV3 semantic segmentation model performed the best, with a pixel-level IoU of 0.780, an object-level precision of 89.3% and object-level recall of 96.9%. Findings suggest that for this data set of images of wooden planks, the benefits of training on labeled data outweigh the time cost of annotation.