Adaptive Retraining Techniques for Foreign Object Detection in Industrial Environments

Foreign-object detection systems based on machine learning perform well when trained on a substantial amount of high-quality data that reflect well the environment in which they will be deployed. In industrial applications, data are often not readily available, and the collection and annotation of data is accompanied by large costs, labour, and time. Moreover, over time, the industrial setting may change, and the previously fine-tuned machine-learning model may not be suitable to the changed setting anymore. We propose a novel active-learning based method for foreign object detection that addresses these problems. Our method strategically selects the samples for automatic labelling or manual annotation on the basis of class-specific accuracy. Moreover, a class-based sampling technique is employed to maintain the class balance to avoid catastrophic forgetting. Experimental results demonstrate that our method achieves comparable accuracy to the model trained with all data with fewer samples, thus reducing cost for labelling, saving time, and lowering computational complexity to retrain.