Overview of Face Recognition Technology

In algorithmic terms, Face Recognition can be defined as the process of mapping a pair of images to produce a similarity value (range of equivalence) for the similarity between the paired images.

In other words, when verifying a given input face photo with a provided template face photo, a score will be produced to represent the similarity between the two face photos (which is then called the similarity value).

In the face recognition process, there are two contexts for the matching process: verification and identification. If face recognition is used for verification, it means that the process of matching a face photo with a single reference photo is performed (1:1). On the other hand, for the identification process, face recognition will work by querying a gallery containing many photos (1:N) to produce the top-scoring possible faces.

The performance of a face recognition system depends on the quality and conditions of the identified or verified face photos. Therefore, the similarity value (similarity value) has the potential to be different.

To determine whether a face is considered to be a match with the reference database, we need to determine the "match" (match) and "non-match" (no match) similarity value limits, which are then called the threshold. By setting a threshold, it means that we can "tweak" the face recognition system to determine the pass/fail score matching limit as desired, but based on accurate measurements. For example, the threshold is set at 80%, so that two photos with a similarity value above 80% will be considered the same person.

Users of face recognition services generally have recommended threshold values, but the actual threshold value can be changed according to needs.

A high threshold value will produce fewer cases in the match category, while a lower threshold will produce more cases in the match category. Is a very high threshold value better? In reality, this may not be the case, but a higher threshold will be more selective. On the other hand, a threshold that is too high may reject cases that are actually matches or produce false rejects.

Last updated