Performance Comparison of Soiling Detection Using Anomaly Detection Methodology

초록

The anomaly detection task is widely exploited in the industrial environment for identifying defective parts in the process of manufacturing. This paper extends anomaly detection task to detect distinct types of soiling patterns in vehicle camera lens. For both training and testing, input images are resized and cropped, and soiling masks are converted to white color regardless of pre-annotated regions. In the performance evaluation, the patch- wise detection methods outperformed the reconstructive approach and the probabilistic approach by 1.6% and 0.7% AUROC, respectively, in anomaly detection task.

키워드