Back to results
US 12,591,646 B2Granted
Multimodal biometric fusion based authentication
AssigneeThales Dis France SAS
InventorsYi Ding, Daniel Asraf, Jing Jim Wang
Life of this patent
Estimating the term…
Abstract
Provided is a method for recognizing an individual by fusing a first match result generated by a first single-modal biometric recognition system of a first biometric modality and a second match result generated by at least one second single-modal biometric recognition system of a second biometric modality into a fused match result. Machine leaning based on training of retrieved data is performed to obtain optimized parameters meeting a target performance metric. Parameters of a fusion model are updated with the optimized parameters in a database, which is then applied to match results to recognize the individual. Other embodiments disclosed.
Claims18
A method for recognizing an individual by fusing a first match result (MatSco1;
- RanLisi) generated by a first single-modal biometric recognition system (SModSys1) of a first biometric modality (BMod1) and at least one second match result (MatSco2, . . . , MatScoN;
- RanLis2, . . . , RanLisN) generated by at least one second single-modal biometric recognition system (SModSys2, . . . , SModSysN) of a second biometric modality (BMod2, . . . , BModN) into a fused match result (FSco;
- FRan), the method being suitable to be implemented by a multimodal biometric fusion system (FSys) and said method comprising:
- by a training engine (TE):
- receiving a training request (TReq) for optimizing parameters of a fusion model (FM) to achieve a target performance metric (FARt);
- retrieving, from a database (DB) of the fusion system, the fusion model (FM), initial parameters (ParInt) of the fusion model (FM), a training sample set (TSet1) of the first biometric modality (BMod1) and at least one training sample set (TSet2, . . . , TSetN) of the at least one second biometric modality (BMod2, . . . , BModN);
- performing a machine learning based training on the retrieved data (FM, ParInt, TSet1, TSet2, . . . , TSetN) to obtain optimized parameters (ParOpt) meeting the target performance metric (FARt);
- updating the parameters of the fusion model (FMt) with the optimized parameters (ParOpt) in the database (DB);
- by a fusion engine (FE) of the fusion system:
- receiving, from a fusion controller of the fusion system, a message (M) comprising a fusion request (FReq), the first match result (MatSco1;
- RanLis1) generated by a first single-modal biometric recognition system (SModSvsI) of a first biometric modality (BModI), and the at least one second match result (MatSco2, . . . , MatScoN;
- RanLis2, . . . , RanLisN) generated by at least one second single-modal biometric recognition system (SModSys2, . . . , SModSvsN) of a second biometric modality (BMod2, . . . , BModN);
- obtaining from the message (M) the first match result (MatSco1;
- RanLisi) and the at least one second match result (MatSco2, . . . , MatScoN;
- RanLis2, . . . , RanLisN);
- retrieving, from the database (DB), the optimized parameters (ParOpt) of the fusion model (FM);
- applying the fusion model (FM), with the optimized parameters (ParOpt), to the first match result (MatSco1;
- RanLisi) and the at least one second match result (MatSco2, . . . , MatScoN;
- RanLis2, . . . , RanLisN) to calculate a fused match result (FSco;
- FRanLis) for recognizing the individual,
- evaluating by an evaluation engine (EE) of the fusion system a performance metric (FAR) based on a set of multimodal biometric fused match results (FSco) that have been generated by the fusion engine (FE), comparing the evaluated performance metric (FAR) with a target performance metric (FARt) to trigger the sending of the training request (TRea) to the training engine (TE).
18 more elements
A multimodal biometric fusion system (FSys) comprising a processor of a training engine (TE) of the fusion system:
- receiving a training request (TReq) for optimizing parameters of a fusion model (FM) to achieve a target performance metric (FARt);
- retrieving, from a database (DB) of the fusion system, the fusion model (FM), initial parameters (ParInt) of the fusion model (FM), a training sample set (TSet1) of the first biometric modality (BMod1) and at least one training sample set (TSet2, . . . , TSetN) of the at least one second biometric modality (BMod2, . . . , BModN);
- performing a machine learning based training on the retrieved data (FM, ParInt, TSet1, TSet2, . . . , TSetN) to obtain optimized parameters (ParOpt) meeting the target performance metric (FARt);
- updating the parameters of the fusion model (FMt) with the optimized parameters (ParOpt) in the database (DB);
- of a fusion engine (FE) of the fusion system:
- receiving, from a fusion controller of the fusion system, a message (M) comprising a fusion request (FReq), the first match result (MatSco1;
- RanLis1) generated by a first single-modal biometric recognition system (SModSys1) of a first biometric modality (BMod1), and at least one second match result (MatSco2, . . . , MatScoN;
- RanLis2, . . . , RanLisN) generated by at least one second single-modal biometric recognition system (SModSys2, . . . , SModSysN) of a second biometric modality (BMod2, . . . , BModN);
- obtaining from the message (M) the first match result (MatSco1;
- RanLis1) and the at least one second match result (MatSco2, . . . , MatScoN;
- RanLis2, . . . , RanLisN);
- retrieving, from the database (DB), the optimized parameters (ParOpt) of the fusion model (FM);
- applying the fusion model (FM), with the optimized parameters (ParOpt), to the first match result (MatSco1;
- RanLis1) and the at least one second match result (MatSco2, . . . , MatScoN;
- RanLis2, . . . , RanLisN) to calculate a fused match result (FSco;
- FRanLis) for recognizing the individual,
- evaluating by an evaluation engine (EE) of the fusion system a performance metric (FAR) based on a set of multimodal biometric fused match results (FSco) that have been generated by the fusion engine (FE),
- comparing the evaluated performance metric (FAR) with a target performance metric (FARt) to trigger the sending of the training request (TReq) to the training engine (TE).
15 more elements
Dependencies are inferred from the claim text.