Some remarks on selected image analysis problems using multivariate statistical methods

The task of making inferences from a digital image frequently revolves around the ability to use multi-dimensional data or feature vectors optimally. This paper proposes directions in the handling of feature vectors with multivariate statistical methods with illustrations from three areas of applications. Experience shows that wherever possible, the lowest dimension of the feature vector is preferred since this increases the chance of deriving appropriate probability distributions for optimal inference. When the dimension of the feature vector is large, dimension reducing techniques should be considered. This paper shows that optimal statistical inference may be achieved if dimensions, probability distributions, relationship between variables and possibly outliers are simultaneously considered.