Automated visual inspection of manufactured parts using deep convolutional neural networks and transfer learning

Abstract

Most manufacturing processes involve some form of visual quality control of the produced parts. Automated solutions can reduce the required manual work significantly while increasing reliability. However, common obstacles to the construction of smart visual inspection systems are the complexity of the inspected parts, varying types of defects, and small datasets. In this study, we apply state-of-the-art convolutional neural networks to classify infrared images of thermal conductive components manufactured in a real factory setting. Typically, training deep neural architectures requires very large datasets, but this effect is mitigated by using transfer learning. The dataset consists of 6,000 images with 4,200 defect samples and 1,800 intact samples, including different types of flaws and component models. We present a concept for implementing the automated visual inspection system, including dataset preparation, model training, and the inline application. The goal is to establish a Human-in-the-Loop approach, that maximizes accuracy and safety while keeping the required human work at a minimum. A key finding of our research is that dataset preparation and cleaning had a greater impact on the classification accuracy than the optimal choice of the model or training parameters.