Method and system for anomaly detection in data sets

Abstract

A method and apparatus for detecting and segmenting anomalous data in an input data set such as an image is described, which makes use of a normalized distance measure referred to as a zeta distance score. A test data point from an input test data set is compared with its corresponding nearest neighboring standard data points in standard data sets representing variation in normal or expected data values, and the average distance from the test data point to the standard data points is found. An additional average distance measure representing the average distance between the different nearest neighboring corresponding standard data points is also found, and a normalized distance measure obtained by finding the difference between the average distance from the test data point to the standard points and the average distance between the nearest neighboring standard data points themselves. Where the input test data set is an image then a zeta distance score map can be found. By then thresholding the zeta distance scores obtained for the input data set using an appropriate threshold, anomalous data in the data set with a high zeta distance score can be identified, and segmented.