Extended subspace methods of pattern recognition

The Subspace Pattern Recognition Method (SPRM) is a statistical method where each class is represented by a separate subspace. There are a number of variants to it including the Averaged Learning Subspace Method (ALSM). The decision surfaces in all these methods are quadratic. In some applications, we may require decision surfaces which are more nonlinear in nature. In this paper, we have proposed the use of more than one subspace (cluster) in the representation of the classes in the subspace methods of pattern recognition. By keeping the number of principal components in all the clusters the same, this model allows for a piecewise linear approximation of the decision surfaces. We have used this model to extend both the SPRM and the ALSM to obtain the Extended SPRM and the Extended ALSM, respectively. We have investigated the use of a dynamic clustering algorithm for the assignment of patterns to different clusters as opposed to a random assignment. We have conducted experiments on three data sets including a 192-dimensional large character data set. The results indicate that the proposed methods have the potential to approximate any decision surface, and can considerably improve the classification accuracy on the test sets.

[1]  Josef Kittler,et al.  Pattern Recognition Theory and Applications , 1987, NATO ASI Series.

[2]  A. Rosenfeld.  Image pattern recognition , 1981, Proceedings of the IEEE.

[3]  King-Sun Fu,et al.  Digital pattern recognition , 1976, Communication and cybernetics.

[4]  M. Narasimha Murty,et al.  Hebbian learning subspace method: A new approach , 1997, Pattern Recognit..

[5]  Jack Sklansky,et al.  Locally Trained Piecewise Linear Classifiers , 1980, IEEE Transactions on Pattern Analysis and Machine Intelligence.

[6]  Erkki Oja,et al.  Texture subspaces , 1987 .

[7]  Bernd Fritzke,et al.  Growing cell structures--A self-organizing network for unsupervised and supervised learning , 1994, Neural Networks.

[8]  Erkki Oja,et al.  Subspace methods of pattern recognition , 1983 .

[9]  Chong-Ho Choi,et al.  Constructive neural networks with piecewise interpolation capabilities for function approximations , 1994, IEEE Trans. Neural Networks.

[10]  Terrence J. Sejnowski,et al.  Learned classification of sonar targets using a massively parallel network , 1988, IEEE Trans. Acoust. Speech Signal Process..

[11]  Alireza Khotanzad,et al.  Rotation invariant image recognition using features selected via a systematic method , 1990, Pattern Recognit..

[12]  ERKKI OJA,et al.  The ALSM algorithm - an improved subspace method of classification , 1983, Pattern Recognit..

[13]  Satosi Watanabe,et al.  Knowing and guessing , 1969 .

[14]  M. Narasimha Murty,et al.  A genetic approach for selection of (near-) optimal subsets of principal components for discrimination , 1995, Pattern Recognit. Lett..