Feature learning and clustering analysis for images classification

The problem this thesis is addressing is to improve an existing classification in 10 categories of the images captured by SEM microscopes. In particular, the challenge faced is to classify those images according to a hierarchical tree structure of sub-categories without requiring any further human labelling effort. In order to uncover intrinsic structures among the images, a procedure involving supervised and unsupervised feature learning, as well as cluster analysis is defined. Moreover, to reduce the bias introduced in the supervised phase, various strategies focusing on features of different nature and level of abstraction are analyzed.