Robust Foreign Object Detection in Aqueous Food Media using Circumferential-averaged Surface Electric Potential and Neural Networks

Current approaches for detecting foreign objects (FOs) in the food production process are mostly based on mechanical separation or imaging techniques that only detect specific types of FOs, which may have limited applicability for detecting FOs of different conductivities, shapes, and sizes. Previous work has shown that convolutional neural network (CNN) classifiers can reliably detect and classify FOs based on simulated surface potential image data, but the performance can be susceptible to high noise interference. This paper introduces a straightforward approach to transform surface potential image data into one-dimensional circumferential-averaged signal data, which effectively reduces data dimensionality and noise variance in signals, while preserving important spatial signal characteristics. Evaluation on unseen simulated data and real data acquired in a controlled laboratory environment demonstrates that the circumferential averaging technique yields promising performance in both FO detection and classification tasks. This suggests that the new lightweight approach could have a significant advantage for fast and reliable FO detection in food processing plants.