Facial recognition

I. Overview of Facial recognition

Facial Recognition Technology is a robust technology which has ability to detect and recognize a person whose face image stored in the database.This technology has been researched for a long time but the used technology had several limits including low accuracy and low speed of prediction. However, in this era of big data and Deep learning, Facial Recognition Technology reached a new peak. Acknowledging the huge potentials this technology could bring to human’s life, leading nations in technology have invested time and effort in digging deep into this field. An example is China with AI applications in citizen management and criminal detection or United State with several applications applied in security and other areas.

II. Technique

We have researched a variety of state-of-the-art Deep Learning models and their variants. After trying those different models, we chose to use feature extraction in combination with One-Shot Learning, which gave the result having highest accuracy.

1. Feature Extraction

 

Feature extraction by CNN is a modern image processing method based on deep learning. It has the ability to detect and extract specialized features based on the color and shape of the processing digital image through pooling windows. Because computer cannot process image in the same way as human, feature extraction enables the computer to have an overview of the photo, that result in higher accuracy when predicting.

2. One-shot learning

One-shot learning is a technique to classify objects. Being different from using a label to each object, this method encodes an object into a multidimensional vector. Since every person have an individual label, labeling each new photo like the conventional practice would required the model to be updated repeatedly, making the model structure become too large and hard to maintain. Therefore, one-shot learning seems more useful and effective than labeling photo individually. The question here is: How can we create characteristic vector of each face. This problem is illustrated in the picture below. Assume that a person A is taken photo at other scene having picture P, and we have a different person called N. Computer has to learn how to encode the value of A and P more similar than value of A and N. After this network is trained with more than 1,000,000 people, the network with be good enough to distinguish different faces.

 

III. Applications of Facial recognition

  • Criminal Identification: System of cameras equipped with facial recognition technology can be used to identify criminals in the airport, train station or other public areas. It also can be helpful to keep track of a person in large areas.
  • Access and Security: Facial biometrics can be integrated with physical devices to replace traditional type of password. It also can be applied in check-in, check-out system of companies or in smart door lock of households.
  • Customer statistics: Valuable facial analytics information of customers thanks to facial recognition in combination with age and gender detection help retailers to serve customers better.
  • Find missing people: Facial recognition tool can help a missing child meet his/her parents by rapidly and accurately matching pictures with their biological parents.  

IV. Advantages of current technology

  • Accuracy higher than 98%, it is implemented in our system for check-in and check-out
  • Required only low spec hardware.
  • Can be integrated in diverse platforms: IoT, mobile (IOS, Android, Windows).
  • Easy to use for end users.
  • Have ability to detect multiple faces simultaneously in a picture, video.
  • Predict real time