Abstract

Support Vector Machines have been used successfully for the classification of data in a wide range of applications. A key factor affecting the accuracy of the classification is the choice of kernel. In this paper we propose the use of Support Vector Machines with a correlation kernel. The correlation kernel is an appropriate choice when performing classification of Raman spectra because it reduces the need for pre-processing. Pre-processing can greatly affect the accuracy of the results because it introduces user bias and over-fitting effects. The correlation kernel is “self-normalizing” and produces superior classification performance with minimal pre-processing. Our results show that the performance on highly-noisy data, obtained using inexpensive equipment, is still high even when the classification is applied on a distinct hold-out set of test data. This is an important consideration when developing clinically viable diagnostic applications.

© 2011 OSA/SPIE