A0272
Title: Model based labelling of hyperspectral food images
Authors: Ganesh Babu - University College Dublin (Ireland) [presenting]
Abstract: Food engineers have been using hyperspectral images to classify the type and quality of a food sample under study, typically using machine learning (ML) classifiers. In order to train these classifiers, threshold-based approaches are used to label the pixels and classifiers are then trained based on those labels. However, the threshold-based approaches are ad-hoc, subjective and cannot be generalized across hyperspectral images. To address this issue, a consensus-constrained parsimonious Gaussian mixture model (ccPGMM) is proposed to cluster the pixels of the hyperspectral images. The resulting cluster labels can then be employed for training the classifiers in the classification phase. The ccPGMM utilizes visual and spatial information from appropriately selected subsets of pixels. This information is used as a constraint when clustering the rest of the pixels in the image. In this way, the model ensures that pixels in the same areas of the image are clustered together, while pixels located in different regions of the image are not allocated to the same cluster. The hyperspectral images are high-dimensional - a parsimonious latent variable model and a consensus clustering approach are employed to handle this. With the consensus approach, the data are divided into multiple subsets of variables and the constrained clustering is applied to each subset. The clustering results are then combined across all the subsets to provide a consensus clustering solution for each pixel.
