Leveraging hyperspectral quality inspection in the agrofood sector with AI

Hyperspectral imaging allows capturing the continuous spectrum of light at each pixel in the visible and near-infrared bands. It provides more optical information than common RGB images and the human eye. As a result, it has found success in industrial applications in waste and food sorting quality control. Recent deep learning-based image processing algorithms outperform classical computer vision techniques in these kind of tasks. However, these algorithms require large amounts of labeled training data, which are not as extensive and readily available for hyperspectral images as for common RGB datasets. Therefore, efficient processing and training techniques need to be developed to successfully exploit the abundant amount of information provided by hyperspectral imagining. The focus of this PhD relates to the development of a digital twin for spectral cameras, calibration transfer through simulated images from the digital twin and the implementation and benchmarking of spatial-spectral AI algorithms.