Quick Match: Fast Template Selection for Fingerprint Recognition

Sammanfattning

Fingerprint recognition has become a popular authentication method in modern smartphones, replacing traditional passwords and PINs. However, smartphone fingerprint sensors are small and capture only partial fingerprints with limited resolution. Combined with varying image quality and inconsistent finger placement, multiple images must be stored per user to ensure reliable authentication. During verification, the scanned fingerprint must be compared against all of these images. Since these comparisons are computationally expensive, a fast sorting algorithm is used to rank the stored images and forward only the most promising candidates to the matching algorithm. The current sorting method uses histogram-based representations of fingerprint features, but whether this is the most effective approach remains an open question. This thesis investigates four approaches to improve the sorting algorithm, looking into the trade-off between sorting performance and computational time. Double Angle builds upon the current method, but uses ridge orientation symmetry to reduce computational time by a factor of four. Curvature replaces an existing feature with ridge curvature measurements. Overlapping Regions partitions images into subregions and employs machine learning models to combine comparison scores. Embeddings uses residual networks trained with triplet loss to learn compact representations for cosine-similarity comparison, testing both conventional and rotation-equivariant architectures. Experiments on optical and ultrasound sensor databases evaluated how effectively each approach retained true matches among top-ranked templates. Double Angle improved sorting performance by 1.09 percentage points on average while reducing computational time by a factor of four. The Embeddings approach achieved the strongest results, improving performance by 7.99 percentage points on real-world data using the conventional architecture. Curvature decreased performance, while Overlapping Regions only showed limited improvements on certain datasets, despite significantly increasing computational time.