Dorsal finger texture recognition: Investigating fixed-length SURF

We seek to create fixed-length features from dorsal finger skin images extracted by the SURF interest point detector to combine it in the privacy enhancing helper data scheme. The source of the biometric samples is the GUC45 database which features finger vein, fingerprint and dorsal finger skin images for modality fusion. First, the region of interest (ROI) is extracted, after which SURF features are extracted, and finally two different approaches for creating fixed length feature vectors are applied. SURF performance on the ROI is comparable to the PolyU database reported in the literature, namely an equal error rate of 0.74%. Of the two explored approaches for fixed-length features creation, averaging the descriptor components proved the most successful, achieving an equal error rate of 11.72%. Potential run-time performance increases were discovered as a side-effect. Without changing the complexity of the SURF matching scheme, a reduction in run-time of 75%–80% has been achieved, with only minimal precision loss; EER increases from 0.74% to 1%. The complexity of the matching can be reduced from O(n²) to constant time, but at a higher precision cost and resulting in an EER of 16.51%.