Active Learning with Continuous Synthetic Data Generation for Industrial Quality Inspection – Jacob Henningsson

Master's Thesis presentation by Jacob Henningsson, Master's programme in Image Analysis and Machine Learning

Quality assurance is an important aspect of manufacturing. However this process can be error prone and resource intensive when preformed by humans. Automating quality assurance with deep learning vision models is an enticing proposition, but suffers from limited data availability. Generating synthetic data is an alternative to classical data collection. However, solutions using synthetic data often underperforme in specific classes and settings. To reduce underperformance, this thesis proposes to generate data during training, based on model evaluation on difficult data samples. A combined data generation and training pipeline for training deep learning models has been created. Information of model performance is gathered during training and is used to update the configuration parameters for dataset generation. The effectiveness of continuous data generation during training is explored in an ablation study. We find continuous generation to be generally beneficial over static dataset training.