72‐3: Deep Learning Based Visual Defect Detection in Noisy and Imbalanced Data

Authors:

Abstract
Despite the immense potential, deep learning‐based visual defect detection is often impaired by the inadequate quality of real‐world data. Here, we leverage Confident Learning and One‐class Classification to tackle two major data issues: 1) mislabeling and 2) limited defects for model development, achieving ≥0.85 AUROC on intentionally polluted industrial data.