Hierarchical Clustering from ICA Mixtures - On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling - page 95

Fig. 4.6 Hierarchical classification from ten ICA mixtures (Laplacian source distributions).
a Data, b dendrogram estimated by the proposed method, c dendrogram estimated by the single-
linkage algorithm
have been related with the detection of borders in natural images [ 26 ]. Therefore,
basis functions have a physical relation with objects and can be used to measure the
similarity between objects based on ICA decomposition. In image patch decompo-
sition, the set of independent components is larger than what can be estimated at one
time, and the result at one time is an arbitrarily chosen subset [ 23 ]. Nevertheless, ICA
has been applied successfully in several image applications; see for instance [ 29 ].
Neurophysiological primary visual cortex activities have also been related with the
detection of borders in natural images. This has allowed theoretical dynamic models
of the abstraction process in living beings to be proposed (which proceeds from the
visual cortex to higher-level abstraction) [ 30 ].
This section is intended to process the parameters of ICA mixtures in order to
obtain hierarchical structures from the basis function level (edges) to higher levels
of clustering. Specifically, the proposed clustering algorithm is applied to image
analysis obtaining promising results in discerning object similarity (object rec-
ognition task) and suitable levels of hierarchies by processing patches of natural
images. Thus, the feedforward process that is featured by the hierarchical clus-
tering method can be related with the process of abstraction.
4.5.1 Real Object Recognition
For the object recognition task, the Columbia Object Image Library -100 (COIL-100)
database was used [ 31 ]. The database consists of a large number of images of objects
over a black background (views taken from different angles). The size of the images
is 128 9 128 pixels. Some of the object images are very similar to others depending
on the angle from which the picture of the object was made.