A Novel Out-of-Distribution Detector based on Autoencoder and Binary Classifier with Auxiliary Input
Sung-Min Kim, Kyoung-Hyoun Kim · Journal of Student Research · 2022
At industrial production levels, anomaly detection is crucial to maintaining adequate levels of safety standards and quality assurance in products. Numerous previous research focuses on detecting the anomaly samples yet their methods cannot filter the unexpected outliers that have completely different image features. In this research, we propose a novel out-of-distribution detector based on an autoencoder and binary classifier with auxiliary input. Given the input image, the autoencoder produces the latent variable and reconstructed image. The difference image is generated and fed to the binary classifier to classify whether the input image is an outlier or a normal sample. The latent variable which contains useful feature-level information is fed to the intermediate layer of the classifier to produce the precise classification results. We also propose noise-addition augmentation to make the trained model consistently perform against various kinds of noise-containing images which are often found in real-world scenarios in industrial environments. The proposed method achieves AUC of 0.9718 and 0.5908 in the MNIST and CIFAR10 datasets, respectively. Through these experiments, we have shown that the proposed method outperforms the previous state-of-the-art methods.