On Improving Performance of Surface Inspection Systems by On-line Active Learning and Flexible Classifier Updates

Classification of detected events is a central component in state-of-the-art surface inspection systems that still relies on manual parametrization. While machine-learned classifiers promise supreme accuracy, their reliability depends on complete and correct annotation of an extensive training database, leaving the risk of unpredictable behavior in changing production environments. We propose an active learning-based training framework, which selectively presents questionable events for user annotation and is capable of online operation. Evaluation results on two data streams from microfluidic chips and elevator sheaves production show that annotation effort can be reduced by 90% with negligible loss of accuracy. Simulation runs introducing new event classes show that the on-line active learning procedure is both efficient in terms of learning speed and robust in maintaining the accuracy levels of existing classes. The results underline the feasibility and potential of our approach that significantly reduces the required effort for inspection system setup and adapts to changes in the production process.