The task of a classifier is to determine the best class for an incoming feature vector.

Each person enrolled in BioID is assigned a unique class, and the classifier compares a new recording (i.e. the feature vectors that are extracted out of this recording) with all (formerly trained and stored) prototypes of each class.

The prototype with the highest similarity determines the class ("winner takes all" principle).

The output of a classifier therefore is the Class ID (the person with the best matching features) - the similarity value. Similarity values are typically in the range between 0.0 and 1.0, a match with 1.0 would mean a perfect match. Note that values in biometrics are never really 100%.

That means:

Since BioID uses individual classification mechanisms for each biometric channel, each channel has its own threshold. Thresholds can be altered in the BioID User Interface within reasonable limits (the system doesn’t allow to set ridiculous threshold values like 0.0 or 1.0).

Sensor Fusion

For the analysis of the classification results, BioID combines the output of multiple classifiers in such a way that it can obtain different security levels.

First of all, each single trait used must have a classifier (similarity) value exceeding a threshold specifically designed for this classifier. Only if all traits have higher similarity values than their pre-adjusted thresholds, the overall result is investigated further. If one of the traits fails, the whole classification attempt will be rejected. By this approach, BioID can hardly be faked with e.g. just a photograph and a tape of a voice recording - as long as also the lip movement is active and is not within reasonable similarity, the attempt to be recognized will be rejected.

After the single-threshold test, all classification results are summed. Hereby the individual traits can be weighted differently. When, for example, lip movement always reliably recognizes a person, this feature can be taken more into account than others in the classification. The weighted sum must exceed the overall threshold (which can be set by BioID’s security slider) - only then the person will be recognized.

BioID is pre-set with an average security setting. To get a feeling about what happens when you change the security value (i.e. change the overall threshold), please realize the dependency of false acceptance rates (FAR) and false rejection rates (FRR) as described in About FAR, FRR and EER.

As described in the section BioID Multimodality, the fusion of single classification results significantly increases the classification performance and accuracy of a biometric recognition system.

Multimodal Accuracy

To demonstrate the effect of BioID's fusion technology, some results of an internal test are presented. The test "AV1" consisted of 22 people having 10 recordings each. Low-budget webcams and microphones were used. Face, voice and mimic features were extracted and classified. The purpose of the test was not to yield excellent classification rates, but instead to look at the performance under bad conditions (such as using different microphones and cameras, under different lighting conditions). Here are the results for the single modalities (if you have problems in understanding the graphs, refer to section About FAR, FRR and EER):

Face

Voice

Mimic

The results above have been calculated by doing verification. In this scenario, one can easily see that the equal error rates range from 3.9% up to 13.1 % (in this case because of the slow cameras used).
 
Now here is the combined result of BioID's fusion of the face, voice and mimic features:

Fusion

The combination does not only improve the EER significantly to 0.7%, by choosing the appropriate thresholds, also each FAR/FRR combination is much better than every single modality.