The theory of advanced correlation filters has evolved from the literature of optical pattern recognition in the last two decades; they have proved effective classifiers in a number of applications, among them biometric recognition and automatic target recognition. Correlation filter designs use the image intensity domain of training examples to compute a class template that produces characteristic correlation outputs to distinguish between authentic users and impostors. When applying the filter for testing the authenticity of a new target image, the output plane is expected to have a shape containing a correlation peak if the image is authentic, but no such peak if the image belongs to another class. Properties of correlation filter classifiers include graceful degradation, shift invariance and closed-form solutions.
The code has been tested using fingerprint images taken with an UPEK swipe fingerprint reader with capacitive sensor and USB 2.0 connection. Database is 16 fingers wide and 8 impressions per finger deep (128 fingerprints in all). We have obtained the following results:
One-to-many fingerprint identification: using 2 images for each finger randomly selected for training and the remaining 6 images for testing (totally 32 images for training and 96 images for testing), without any overlapping, we have obtained an error rate smaller than 0.6% (top one error rate).
One-to-one fingerprint verification: we have obtained an EER equal to 5.6641%.
Index Terms: Matlab, source, code, correlation, filters, AFIS, automated, fingerprint, identification, system.
Windows Server 2008