Semi Supervised Anomaly Detection Using PatchCore Auto Encoder
P G Abhishek, Deepan Raj M, G. Sainarayanan · 2023
Anomaly detection is always a challenging problem in the field of computer vision because the number of defective samples is often minimal or nonexistent. This creates a significant challenge in understanding the nuances of an anomalous sample without having seen it. This often remains a cold-start problem as most approaches don't consider anomaly samples, even if they are present in small numbers. The proposed method is an extension of patchcore-based anomaly detection, by taking anomaly samples into consideration. The method extracts locally aware patch features from nominal samples and is stored in a memory bank and then furthermore training the features with an autoencoder to learn the nuances required for reconstruction. Additionally, abnormal samples are included in the process, and stored in the same memory bank. Finally, using the K-Nearest Neighbor, the test samples can be classified there by auto encoder will be used to point out the region of interest for the likelihood of anomaly. The weak supervision of abnormal data samples and autoencoder greatly enhances the significance of the work and can achieve an accuracy of 94% on the MVTEC AD Benchmark (Pill) dataset.