Polar-Ano: Surface Anomaly Detection via Deep Polarization Imaging and Data Synthesis with Physic-based Rendering

Yaxu Xie, Alain Pagani, Didier Stricker

Abstract: In this paper, we propose a complete pipeline from generating polarization images via physics-based rendering to train and deploy an image anomaly detection and localization model for polarimetric industrial inspection. The method consists of two stages. We first compute the Polarimetric Priors with both determined and learning-based method. Then, the Polarimetric Priors are given to a self-supervised surface anomaly detection network to predict the anomalies score and anomalies masks. To train the network, we adapt and modify a physic-based rendering pipeline to generate photo-realistic data samples of polarized images on a large scale. Our experiments show the effectiveness of our proposed pipeline.