An Anomaly Detection System for Transparent Objects Using Polarized-Image Fusion Technique

Abstract

An anomaly detection system using a polarized-image fusion technique has been developed for food inspection applications. It is capable of detecting (a) foreign objects among objects wrapped in transparent reflective material and (b) transparent foreign objects in transparent bottles. The conventional anomaly detection system using a traditional RGB camera has low accuracy for such detection, due to the large amount of glare that can occur from reflective surfaces. Regions with glare are often falsely perceived as anomalies. Since transparent foreign objects have few features, they are difficult to recognize. To address these problems, a polarized-image fusion (PIF) technique is developed. Four polarized images are fused to synthesize a high-quality image where glare is suppressed, and transparent foreign objects are highlighted. These polarized images are captured simultaneously by a single camera utilizing an advanced polarized CMOS image sensor. The PIF technique was evaluated with two kinds of data set: (1) cookie samples wrapped in transparent plastic bags and (2) transparent plastic bottles containing transparent plastic foreign objects. High anomaly detection accuracies of 0.851 AUC (area under receiver operating characteristic curve) for the cookie sample data set and 0.871 AUC for the plastic bottle data set were achieved. Compared with the deep one-class classification neural network with simple RGB data input, the accuracies were improved by 0.09 AUC for both cases.