Improved Anomaly Detection based on Loss Prediction
Wenkang Zhang, Fengqian Pang · 2022 4th International Conference on Communications, Information System and Computer Engineering (CISCE) · 2022
In the field of anomaly detection, it is challenging to collect and annotate enough samples of anomalies. Unsupervised learning is therefore frequently employed in anomaly detection and localization. The existing deep learning models for anomaly detection leverage various loss functions to train the neural networks, and they have achieved excellent performance. However, these methods are only aware of the function that loss values serve in back-propagation, but ignore that the loss value itself can partially represent how anomalous the input sample is. Consequently, we proposed a novel framework that constructs a network branch for loss prediction based on an anomaly detection network, i.e. the base model. This loss prediction branch predicts the loss value in forward-propagation, and the predicted loss value is further integrated with the output of the base model for more accurate anomaly detection. On the MVTec AD dataset, the experimental results reveal that the proposed framework outperforms other mainstreaming methods. The proposed framework can achieve AUC metrics of 98.1 % at the pixel level and 99.5% at the image level, which are 0.2% and 0.4 % higher than the base model, respectively.