A systematic and reliable approach to pattern classification

A systematic and reliable approach to classify patterns is proposed when no a priori information except a set of pre-classified data is provided. A classifier is selected from a number of state of the art pattern classification schemes which are diverse in approach as well as the assumptions employed in their design. The selected schemes include the k-nearest neighbour classifier (kNNC), the minimum Mahalanobis distance classifier (MMDC), and the artificial neural network classifier (ANNC). In order to ensure that the selected classification scheme is properly designed and correctly implemented, the given pre-classified data is analysed, and the relative performance of the classifiers are cross validated as well as compared with a benchmark performance measure. The given data set is subjected to data validation, data visualization and feature quality analysis with a view to detect bad data, to obtain a qualitative picture of the class separability, and to derive a benchmark performance measure called the Bhattacharyya distance measure. In the design phase, the classifiers are executed in the order of increasing accuracy and increasing complexity so that a classifier at one level in the hierarchy sets the performance goal (e.g. classification accuracy) for the task at the next level. Further, to ensure a peak performance, the classifier accuracy is compared with the Bhattacharyya distance measure. The proposed scheme is evaluated on both simulated as well as actual data obtained from the images of the biological cells.