Defect detection in manufacturing quality control using Faster R-CNN

Manufacturing companies spend significant amount of money and employ several subject-matter experts for their quality control processes. Partially or fully automating these processes using machine learning would lead to significant cost and time saving. This is a feasible opportunity also thanks to the continous development of new machine learning models and to the affordability of cloud services like Azure Machine Learning. This study has the purpose of improving the defect detection deep learning model currently used in production for the visual quality inspection of industrial products. The proposed work explains how the performance can be significantly improved (from 0.20 IoU to 0.60 IoU) by using a more balanced dataset including both detective and non-defective samples and performing an accurate hyper-parameter tuning. The results suggest that the Faster R-CNN model is the best performing neural network on this dataset, confirming its capability to recognize small objects in images. A visual inspection of the dataset however suggests that the data collection process could be further improved by employing different data annotators and relying on high-quality cameras and lights.