Active Learning in Industrial Defect Detection: A Dynamic Approach to Uncertainty and Representativeness

The storage and annotation of industrial data is a costly and time-intensive work in developing a defects detector. To overcome this challenge, this study proposes an effective annotation workflow and an active learning strategy that integrates uncertainty and representativeness to construct a reliable defects detection model. The annotation workflow involves both the advisor and the labeling team for the inspection of industrial data that depends on specialized knowledge in a particular area. The active learning strategy combines the uncertainty-based and the representativeness-based sampling metrics flexibly to assess the informativeness of the unlabeled data. Besides, a training strategy for active learning models is used to address not only data uncertainty but also the generalization of feature distribution across all data. We benchmark the proposed method on Bottled Baijiu Defect Detection Dataset, utilizing RetinaNet as the baseline detector. Results demonstrate that the proposed method significantly reduces labeling costs and data storage requirements, which is crucial for industrial defect detection.