Enhancing Weakly Supervised Defect Detection Through Anomaly-Informed Weighted Training

In numerous real-world situations, acquiring extensive sets of labeled data poses a formidable challenge. This study introduces an innovative approach for enhancing weakly supervised learning (WSL) through anomaly-informed weighted training (WT). The method not only is tested in diverse benchmark datasets such as CIFAR-10 and Fashion-MNIST by simulating a binary classification problem with only prior knowledge of some samples belonging to one class but also is applied to a real-world scenario which specifically aims to detect surface defects, like pitting, on ball screw surfaces. The proposed method leverages anomaly detection techniques to refine the training processes in a WSL setting, thereby effectively addressing the challenge of limited labeled data availability. The results demonstrated significant enhancements in the classification metrics, showing the potential of this method for industrial applications.