Improving fingerprint biometrics by local symmetry and fusion
Fingerprint recognition and biometrics in general have been significantly gaining attention within the last years, not least due to the biometric passport that has been introduced all over Europe and the American immigration management system US-VISIT. The tasks of fingerprint recognition can be broken down to smaller entities, i.e. quality assessment, pre-processing, feature extraction and alignment/matching. Techniques contributing to each of these steps are proposed. Furthermore, several (quality-adaptive) fusion schemes for the combination of multiple fingerprint recognition systems are presented.
Symmetry features are employed in all image processing tasks, i.e. to model the ridge-valley pattern, the minutiae points as well as the singularities of a fingerprint. Furthermore, they allow for an efficient implementation by means of 1D Gaussian filters. An untrained reduced reference approach for image quality assessment is proposed, which enables the localization of impurities and structural shortcomings. In addition to symmetry features, the suggested pre-processing utilizes pyramidal image processing and directional 1D filtering to yield a high-fidelity fingerprint enhancement. The proposed local feature extraction method employs a combination of symmetry features for reliable minutiae detection. After an alignment by means of the minutiae points, a correlation-based matching strategy is suggested to establish correspondence between fingerprints. Although all suggested techniques can be perfectly employed in isolation, a combination is advisable, since it adds to their overall efficiency and practicability (e.g. joint integration in hardware).
Multi-algorithm fusion within the fingerprint modality is exploited as a further resource to increase the joint recognition accuracy, also incorporating the quality of the biometric data. For this purpose we suggest a computationally efficient cascaded fusion scheme and additionally evaluate simple fusion rules and a trained Bayesian scheme.
The practical value of all proposed techniques is corroborated by experiments, involving comparisons to relevant existing methods. Furthermore, the tests were performed on extensive and publicly available databases.