Project - Biometrics-Sensor Fusion
With the evolution of biometric technology, the biometric security protection is becoming more and more reliable. Currently, there are mainly 9 different biometric techniques that catch people's attention, including face, fingerprint, iris, voiceprint, hand signature, , facial thermo grams, hand geometry, hand vein and retinal pattern. In order to build a biometric system that is able to achieve desirable accuracy and operate efficiently, sensor fusion combining 2 or more different biometric approaches may be necessary.
Multi-Biometric Data Acquisition Station (MBDAS)
In order to evaluate the proposed biometric verification/identification algorithms, acquire multiple biometrics from an individual at the same time and test what biometric sensors and sensor resolutions/sizes are reasonable candidates, the MBDAS system is built. Now the MBDAS has the high quality face images and fingerprint images data with some variations. Recently the iris pattern images and voiceprint will be added to the MBDAS.
Face: different lighting conditions including Infrared (IR), with/without glasses and with/without ambient light.
Finger: right index and middle fingers
Weighting + DCA (Distinguishing Component Analysis)
The motivation of weighting + DCA is to find a distinguishing subspace and a set of optimal weighting values to represent a set of signals so that the authentic subject is near the subspace and all the other subjects are far from the subspace. We use this algorithm to perform personal authentication.
For the training stage: the algorithm optimizes some criterion function to build such a subspace and a set of optimal weighting values for each individual.
For the testing stage: The individual claims his/her identity. First, the input data are weighted by the optimal weighting values from the training stage. Then weighted input data are projected onto the individual subspace of the claimed identity. The residue between the weighted data and the reconstructed data is calculated. Finally, the residue is compared with some preset threshold to decide the verification result.
The proposed fusion system of different biometric feature is shown as follows.
The input biometric data are from face and fingerprint image of one subject. We will also integrate the iris pattern into our sensor fusion scheme latter. The feature value of the face, residue, comes from the face verification algorithm, DCA + weighitng. The feature value of the finger, Peak-Side Lobe-Ratio, comes from the fingerprint verification algorithm, correlation filter. And the feature value of the iris pattern, Hamming Distance (HD), comes from the iris verification algorithm 2D Gabor wavelet Decomposition.
The feature values from each modality are fused by the DCA + weighting algorithm proposed.We have do experiments using this approach based on our own MBDAS database. Experimental results show that the proposed scheme outperforms the traditional methods.
Any suggestions or comments are welcome. Please send them to Wende Zhang.