Abstract: Industrial anomaly detection is crucial for quality control and predictive maintenance but is challenging due to limited training data, varied anomaly types, and changing external factors affecting object appearances. Existing methods detect structural anomalies, such as dents and scratches, by relying on multi-scale features of image patches extracted from a deep pre-trained network. Nonetheless, extensive memory or computing requirement hinders their adoption in practice. Furthermore, detecting logical anomalies, such as images with missing or surplus elements, necessitates understanding spatial relationships beyond traditional patch-based methods. Our work focuses on Deep Feature Reconstruction (DFR), which offers a memory- and compute-efficient way of detecting structural anomalies. Moreover, we extend DFR to develop a unified framework for detecting structural and logical anomalies, called ULSAD. Specifically, we improve the training objective of DFR to enhance the capability to detect structural anomalies and introduce an attention-based loss using a global autoencoder-like network for detecting logical anomalies. Empirical results on five benchmark datasets demonstrate the effectiveness of ULSAD in the detection and localization of both structural and logical anomalies compared to eight state-of-the-art approaches. Moreover, an in-depth ablation study showcases the importance of each component in enhancing overall performance. Our code can be accessed here: https://anonymous.4open.science/r/ULSAD-2024.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: Following the reviewers' comments, we have edited our paper. The changes are in pink text. The main change includes:
- We improved the text in Section 3.
- We fixed Table 1.
- In Sections 4.2 and 4.3, we provide additional discussion about the results.
- We added Section 5, discussing the memory and computational complexity of our proposed method.
- We provide limitations of our method in Section 6.
- In Appendix B.1, we added anomaly detection and localization results for MVTecLOCO, focusing separately on logical and structural anomalies.
Assigned Action Editor: ~Yan_Liu1
Submission Number: 2579
Loading