ABSTRACT
Correlation filters (CFs) are a well-known pattern classification approach used in a variety of applications that require localization, tracking, or recognition. A CF is a spatial-frequency array that is specifically synthesized from a set of training patterns to produce a sharp correlation output peak at the location of the best match for a similar (authentic) image comparison and no such peak for a nonsimilar (impostor) image comparison. The underlying premise when using CFs is that this correlation output peak behavior on training data extends to testing data. In this chapter, we discuss the use of CFs for biometric recognition, when in verification scenarios there is limited training data available to represent pattern distortions. Accordingly, the lack of training data can cause the correlation output from an authentic-match comparison to be difficult to distinguish from the correlation output from a non-matching impostor. In this chapter, we will discuss stacked correlation filters (SCFs), a simple and powerful approach to address this problem by training an additional set of classifiers that learn to differentiate correlation outputs from authentic and impostor match pairs. This is done by training a series of stacked modular CFs with each layer refining the output of the previous layer. The basic premise is that because correlation outputs have an expected shape, an additional CF can be trained to recognize such a shape and refine the final output.
