Optimization of score-level biometric data fusion by constraint construction training

This paper illustrates a multibiometric method to optimize the fusion of multiple biometries at the score level. The fused score is a linear combination of the individual scores. As a consequence, well-known traditional linear optimization techniques become suitable to determine the constants to be used in the linear combination. The proposed method uses training to optimize the constants. After experimenting with dummy datasets, a fresh multi-biometric dataset of infrared images has been prepared. The data has been subject to extra distortion and occlusions, and then used to train first the individual biometric systems, based on GoogleNet CNNs, and then the fusion engine. Results obtained through the proposed method have an accuracy over 99% in the best configuration. The system at present performs user verification, but an extension to identification can be obtained by reworking the constraints in the optimization problem. A sketch of such extension is provided.