Hyperspectral image classification using CNN: Application to industrial food packaging

During food tray packaging, some contamination may exist due to the presence of undesired objects. It is essential to detect anomalies during the packaging process in order to discard the faulty tray and avoid human consumption. This study demonstrates the on-line classification feasibility when using hyperspectral imaging systems for real-time food packaging control by using Convolutional Neural Networks (CNN) as a classifier in heat-sealed food trays. A hyperspectral camera is used to capture individual food tray information and fed to a CNN classifier to detect faulty food trays with object contamination. The proposed system is able to detect up to eleven different contamination products (plastic, rubber, etc.). In case of a faulty tray, it is discarded from the automated production line. A specific database with different faults and normal cases was built. Various techniques are proposed for CNN architecture and its training. Results show a global fault detection accuracy higher than 94%. Additionally, the classification process is flexible can be tuned for optimal accuracy detection or FOR (False Omission Rate) improvement to reject any possible fault, depending on the factory requirements. An analysis of different faults is done, showing the limitations for some types of faults. Finally, the algorithm is integrated into the automated control of the manufacturing line, a total computation time between 70 and 105 ms depending on the number of channels selected for each food tray was obtained, which makes it possible to run the manufacturing line at a maximum speed of 14 food trays per second.