ABSTRACT
Image analysis includes a variety of tasks such as image restoration, segmentation, registration, visual tracking, retrieval, texture modelling, classification and sensor fusion. Important application domains are medical imaging, remote sensing and computer vision. Problems involving incomplete data, where part of the data is missing or unobservable, are common, and mixture models can be used in many of these tasks directly or indirectly. The aim may be to recover an original image which is hidden and has to be estimated from a noisy or blurred version (restoration). More generally, the observed and hidden data are not necessarily of the same nature (segmentation). The observations may represent measurements, for example, multidimensional variables recorded at each pixel of an image, while the hidden data could consist of an unknown class assignment to be estimated at each pixel. To give an idea of the variety of uses, mixture models have been used for image restoration (e.g. Niknejad et al., 2015), for image registration (Gerogiannis et al., 2009), visual tracking (Karavasilis et al., 2012), image retrieval (Beecks et al., 2015), texture modelling (Blanchet & Forbes, 2008), classification (Bouveyron et al., 2007) and sensor fusion (Gebru et al., 2016), to name only a few of the relevant papers. However, the most typical and direct use 386relates to image segmentation which can be recast straightforwardly into a clustering task. More generally, in this chapter we will focus on problems that can be posed as labelling or clustering problems in which the solution is a set of labels assigned to image pixels or features.
