Feature Reduction

Image information has already been reduced when constructing the feature vector from primary image features. The reduction is based on heuristic rules about what is necessary to classify the image. Feature reduction, discussed in this chapter, reduces the dimension of feature space. It is based on definitions about redundancy of a feature and its relevance to the classification task. Dimension reduction of feature space is often necessary because the initial information reduction from feature extraction is not based on such kind of definition. Hence, often rather too many features are selected for not removing necessary information before classification. Unsupervised feature reduction reduces feature space based on sample distributions without knowledge of sample labels. Supervised feature reduction requires labeled samples. Examples for the two strategies will be presented preceded by a definition of what is meant by redundancy and relevance into the two classes. Advantages and disadvantages of the different redundancy and relevance definitions will be discussed.