Statistical Extraction of Features from High-Dimensional Processes

Abstract

The special problem associated with pattern recognition of large arrays concerns the selection or extraction of useful discriminatory information. This information, spread more or less evenly over the entire array, must be compressed into relatively few parameters which can then be used as the input vectors for the classifier. Much of the previous work in the field of computerised classification in medicine has been based on an approach in which the diagnostic information has been extracted in terms of amplitudes and positions of individual peaks, in the case of a signal such as the ECG, or in terms of shapes orientations and relative positions of interrelated two-dimensional structures in the case of images. The disadvantage of this type of analysis as the feature extraction stage is the complicated nature of the algorithms used to identify the peaks or structures. Furthermore, they are not only problem specific but they also have to be applied to each new signal or image to be classified.