Gaussian mixture model clustering allows accurate semantic image segmentation of wheat kernels from near-infrared hyperspectral images
Andreas Kartakoullis, Nicola Caporaso, Martin B. Whitworth, Ian D. Fisk
Abstract
In this study, an ad-hoc image processing pipeline has been developed and proposed for the purpose of semantically segmenting wheat kernel data acquired through near-infrared hyperspectral imaging (HSI). The Gaussian Mixture Model (GMM), characterized as a soft clustering method, has been employed for this task, yielding noteworthy results in both kernel and germ segmentation. A comparative analysis was conducted, wherein GMM was compared with two hard clustering methods, hierarchical clustering and k-means, as well as other common clustering algorithms prevalent in food HSI applications. Notably, GMM exhibited the highest accuracy, with a Jaccard index of 0.745, surpassing hierarchical clustering at 0.698 and k-means at 0.652. Furthermore, the spectral variations observed in wheat kernel topology can be used for semantic image segmentation, especially in the context of selecting the germ portion within the wheat kernels. These findings carry practical significance for professionals in the fields of hyperspectral imaging (HSI) and machine vision, particularly for food product quality assessment and real-time inspection. • Hyperspectral imaging high-throughput real-time analysis requires segmentation. • Semantic image segmentation with Gaussian mixture model clustering was applied. • The proposed method allows accurate classification of the embryo region. • Soft clustering algorithms can outperform hard clustering algorithms. • Wheat kernel structural information can be studied through NIR spectral variation.