SEE-OOD: Supervised Exploration for Enhanced Out-of-Distribution Detection

Machine vision systems have been increasingly adopted for defect inspection and classification in manufacturing. Beyond identifying pre-defined defect categories observed during training, these systems are desired to also detect new defect types that are Out-of-Distribution (OoD), which differs significantly from In-Distribution (InD) training data. Existing DNN-based OoD detection methods rely on predictive uncertainty or regularization on limited OoD samples, but they often overfit specific OoD samples and fail to generalize. To overcome these limitations, this paper proposes a novel generative adversarial approach that utilizes limited real OoD samples for supervised OoD sample generation with exploration of OoD space. Our method tackles two major challenges of existing works: first, it provides a supervised OoD generation scheme based on real OoD samples, unlike existing unsupervised methods that generate virtual outliers based on InD data; second, it simultaneously augments OoD samples and explores unseen OoD space, reducing the issues of insufficient real OoD samples and overfitting. The paper provides rigorous theoretical results to demonstrate the efficacy of the proposed approach. Comparison studies with state-of-the-art techniques demonstrate superior generalizability to unseen OoD data. Furthermore, we showcase its practical effectiveness in detecting new surface defects using a real-world 3D point cloud dataset of manufacturing defects.