Method and system for extracting features in a pattern recognition system

Abstract

Method and system for extracting features from measurement signals obtained from real world, physical signals by first forming an invariant component of the measurement signals and then using a technique based on a noise subspace algorithm. This technique first casts or projects the transformed measurement signals into separate subspaces for each extraneous variation or group of variations. The subspaces have minimal over-lap. The recognition of a particular invariant component within a pertinent subspace is then preferably performed using Singular Value Decomposition (SVD) techniques to generate a pattern recognition signal. A series of transformations can be used to form an invariant component called the Scale and Translation Invariant Representation (STIR). In one embodiment, the first step is to form an appropriate time-frequency representation such as the Reduced Interference Distribution (RID) or other distribution whose properties are covariant with translations in time and frequency and changes in scale. A series of additional transformations including a scale transform yield the STIR representation. Features are then extracted from a set of STIR representations taken as examples of the desired signal. The STIR approach removes much of the variation due to translation. In bit-mapped documents, the same translation invariant and scale invariant transformations may be made to regularize characters and words. Also, the same feature selection method functions in an image setting. The method has been found to be particularly useful in word spotting in bitmapped documents corrupted by faxing.