Combine soft and hard biometrics yields best results, research suggests

It is a truism – more data means better results. That is what Spanish scientists found when experimenting with soft biometrics and state-of-the-art, commercial facial recognition algorithms.

The researchers, led by Ester Gonzalez-Sosa of Nokia Bell Labs in Spain, found that soft biometrics – traits that are harder to quantify than fingerprints, for example – can boost the accuracy of facial recognition algorithms.

Soft biometrics is a “valuable complement” in unconstrained scenarios, according to the group’s paper. Artificial intelligence startup DeepAI has posted the research and grouped it with related work.

In fact, the researchers were able to improve relative algorithm performance by 40 percent compared to facial recognition.

The soft biometrics used in the study were age, ethnicity, gender, moustache, glasses and beard. All of the images in the LFW, or labeled faces in the wild, database were manually labeled according to those biometric factors.

In one of two scenarios, the researchers employed a manual estimation of soft biometric modalities. The biometrics were combined with the Face++ (Megvii) and VGG-face facial recognition algorithms.

Researchers paired subsets of the biometric algorithms, and each resulted in improved person recognition. However, combining all six produced the best improvement over facial recognition alone – 40 percent.

In the second scenario, they analyzed the same algorithm components but this time with a set of automatically estimated soft biometrics. Again, there is improvement in recognition, although only 10 percent to 15 percent.

Soft biometrics have also been proposed as a tool for detecting deepfakes.

Article Topics

accuracy  |  biometrics  |  biometrics research  |  facial recognition  |  video analytics