A Deep Metric Learning-Based Anomaly Detection System for Transparent Objects Using Polarized-Image Fusion

Abstract

While visual inspection systems have been widely used in many industries, their use in the food and optical equipment industries has been limited. Transparent and reflective materials are often used in these applications, but existing anomaly detection (AD) systems have low accuracy in their detection due to low visibility. Here, we developed an AD system using a polarization camera for reflective and transparent target objects. Two new techniques are developed. First is the polarized image fusion (PIF) technique which suppresses glare from reflective surfaces while highlighting transparent foreign objects. In PIF, four captured polarized images are fused to synthesize a high-quality image according to calculated weight coefficients. The second new technique is an ArcObj-based deep metric learning technique to improve AD accuracy. The proposed system was evaluated in experiments on three datasets: cookie samples wrapped in transparent plastic bags; transparent plastic bottles; and transparent lenses. High AD accuracies in terms of the area under the receiver operating characteristic curve (AUC) were achieved: 0.88 AUC for the cookie dataset; 0.87 AUC for the bottle dataset; and 0.98 AUC for the lens dataset. Compared to the state-of-the-art AD algorithm (Patchcore), the proposed method improved AD accuracy by 0.09 AUC.