Robust Method for E-Maximization and Hierarchical Clustering of Image Classification



We developed a new semi-supervised EM-like algorithm that is given the set of objects present in each
training image, but does not know which regions correspond to which objects. We have tested the
algorithm on a dataset of 860 hand-labeled color images using only color and texture features, and the
results show that our EM variant is able to break the symmetry in the initial solution. We compared two
different methods of combining different types of abstract regions, one that keeps them independent and
one that intersects them. The intersection method had a higher performance as shown by the ROC curves
in our paper. We extended the EM-variant algorithm to model each object as a Gaussian mixture, and the
EM-variant extension outperforms the original EM-variant on the image data set having generalized
labels. Intersecting abstract regions was the winner in our experiments on combining two different types
of abstract regions. However, one issue is the tiny regions generated after intersection. The problem gets
more serious if more types of abstract regions are applied. Another issue is the correctness of doing so. In
some situations, it may be not appropriate to intersect abstract regions. For example, a line structure
region corresponding to a building will be broken into pieces if intersected with a color region. In future
works, we attack these issues with two phase approach classification problem.