Segmentation of Structural Features in Cheese Micrographs Using Pixel Statistics.

Description of microscopic features is often accomplished by qualitative inspection of micrographs. However, assessing the differences responsible for the variation of rheological properties requires taking measurements of microscopic features. A statistically-founded algorithm is proposed for automatic identification of cheese micro-features in Scanning electron microscopy (SEM) imagery for quantitative analysis. Twenty SEM images of Ragusano cheese were recorded and manually labelled using a different colour coding for each of three features: whey pockets, fat globules, and protein matrix. The marked micrographs were used to train a statistical model. In turn, that model was used to mark the looked-for features on new images of the same type. Namely, pixel probabilities of belonging to the three classes above mentioned were computed, using pixel neighbourhoods at different scales. Each pixel was set within the most probable class. The proposed method could recognize the microstructure of Ragusano cheese in SEM images with higher accuracy (on average, 78.0% correct pixels) than other popular algorithms in the literature (e.g., Maximum Response 8 73.7% correct pixels). Being completely automatic in the operative phase, this algorithm allows to batch-process lots of images with no human intervention. It can also be used for any other application, since it can handle any number of phases.