摘要：Quality control allows companies to verify the products’ conformance to requirements and specifications and thus build customer satisfaction and the brand's reputation. Artificial Intelligence enables higher degrees of visual inspection automation, reducing inspection times while ensuring all products are evaluated under the same criteria. This research addresses the problem of an automated visual inspection setup complemented with a manual revision process to inspect manufactured pieces for which the model is most uncertain. We envision that such a visual inspection process can leverage active learning to continuously improve the classification models’ quality based on newly labeled data. Our supervised classification model achieves an AUC ROC of 0,9894. We use probability calibration techniques to set a quality threshold and decide whether a given product must undergo manual revision or the model's prediction can be trusted. Finally, we explore using Explainable Artificial Intelligence, anomaly maps, and the nearest known labeled image to hint to the user about potential defects and reduce labeling errors while speeding up the labeling process. Considering a 99,90% probability threshold, we found that only 40% of the streamed images would be sent to manual revision. Furthermore, the best quality of manual revision was achieved by hinting the user with the closest known labeled image. Such setting increased the precision, recall, and F1 scores by at least 0,5 points compared to the annotation quality when images were displayed with the original class imbalance and without defect hinting. We performed the experiments on real-world data provided byPhilips Consumer Lifestyle BV.
