Conclusions - On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling - page 174

the method to generalize from close data densities and to detect outliers. This
was compared with the traditional single linkage method that is based on
distance between data objects. Image content similarity between objects based
on ICA basis functions allowed an organization of objects in higher levels of
abstraction to be learned (images of objects with similar shapes were grouped).
Experiments with natural images showed the application to image segmenta-
tion based on the similarity of different patches of the image.
• A new procedure to incorporate sequential dependences in classification of
ICA mixtures has been provided ( Sect. 7.1 ) . The so-called SICAMM method
considers the case of having sequential dependence in the feature observation
record. The algorithm is a sequential Bayes processor, which can be used to
sequentially classify the input feature vector among a given set of possible
classes. A hidden Markov model (HMM) was formulated using estimates of
the class-transition probabilities and ICAMM parameters (mixture matrices,
centroids, and source probability densities). These parameters were estimated
using the proposed Mixca algorithm. Some simulations were presented to
verify the improvement of SICAMM with respect to ICAMM. Moreover, a
real data case was considered: the computation of hypnograms from apnea
patients to help in the diagnosis of sleep disorders. Both simulated and real
data analysis demonstrated the potential interest of including sequential
dependence in the implementation of an ICAMM classifier.
• A pioneer application of the ICAMM methods for NDT using impact-echo
technique has been presented in Sect. 5.3 . The model is intended to defects
with particular shapes or geometries, such as cracks, holes, and multiple
defects. The model defined the determination of the quality condition of
materials inspected by impact-echo (in homogeneous and different kinds of
defective materials) as an ICA mixture problem. The model was formulated
considering the impact-echo overall scenario as a MIMO-LTI system. A class
of defective or homogeneous material was represented by an ICA model
whose parameters were learned from the impact-echo signal spectrum. Thus,
the resonance phenomenon involved in the impact-echo method was taken
into account, and the compressed spectrum composed by contributions of
every channel was formulated as observations for ICAMM. The ICA
parameters of each class defined a kind of particular signature for the dif-
ferent defects.
The proposed procedure was intended to exploit to the maximum the infor-
mation obtained with the cost efficiency of only a single impact. To illustrate
this capability, four levels of classification detail (material condition, kind of
defect, defect orientation, and defect dimension) were defined, with the
lowest level of detail having up to 12 classes. Results from extensive data sets
from 3D finite element models and lab specimens of an aluminium alloy that
contained defects of different shapes and sizes in different locations were
obtained. The performance of the classification by ICA mixtures using Mixca
was compared with LDA and MLP classification. We demonstrated that the