+65 8866 8848 sales@dilu.com.sg
Facial Recognition (“FR”) System Framework

On a high level, Dilu’s facial recognition system is composed of the following 4 modules:

1) Pre-processing

  • Face Detection: identify the location of interested human faces.
  • Normalization: normalize the RGB and depth images
  • 3D Enhancement: Creation of 3D facial models, perform multiple rotations (yaw, pitch and roll) and project the 3D models onto 2D surfaces to achieve multiple RGB and Depth images.

2) Modal Fusion
Modal Fusion at the feature layer using independent branches

3) Feature Enhancement
Introduce attention weights mechanism in the feature maps and enhance the feature of key areas of the face

4) Core Algorithm:

  • DenseNet architecture is deployed for the algorithm
  • Enhance model accuracy with multiple Loss Functions
3D Facial Recognition Algorithm

One of the uniqueness of our facial recognition algorithm compared to many of our competitors is our ability to incorporate 3-dimensional depth information in the recognition process. In fact we are one of the very few AI companies that has developed 3D capabilities, from proprietary software engine to specialized camera modules.

Dilu’s true 3D FR algorithm is based on structured light depth perception and 3D real-time dynamic high-precision reconstruction, it is able to ensure very high recognition accuracy especially in a large-scale (>10m) photo database. This superiority in accuracy has been trusted by clients like Guangzhou Metro which has a base database of over 10 million, and daily ridership of over 8 million.

With a lot more facial feature captured in the enrollment, Dilu’s 3D FR engine is equipped with superior accuracy and anti-spoofing capability, and our 3D structured-light technologies have enabled many clients to adopt facial recognition technologies in immigration checkpoints and FR paymetns which requires superior features that is not possible in conventional 3D facial recognition technologies. 

Dilu 3D FR algorithm is also proven to be more versatile than 2D algorithms. It works well in complex light conditions (backlighting, flashing lights etc.) as the depth channel (infrared based) provides additional information to complement the RBG channel. It also has much higher tolerance on the angle of the face image to up to +/-60 degree in yaw, pitch and roll. Dilu 3D also has excellent results working with partially covered facial images, with up to 1/3 of the face covered.

Dilu 3D FR algorithm is also proven to be more versatile than 2D algorithms. It works well in complex light conditions (backlighting, flashing lights etc.) as the depth channel (infrared based) provides additional information to complement the RBG channel. It also has much higher tolerance on the angle of the face image to up to +/-60 degree in yaw, pitch and roll. Dilu 3D also has excellent results working with partially covered facial images, with up to 1/3 of the face covered.

The following tables summarize the advantages of Dilu 3D FR technology:

 Structurd Light

Dilu’s 3D depth perception technologies is based on Structured Light, and below is a table outlining the comparison vs Time of Flight, another popular algorithm for depth perception.

 

Structured Light Time of Flight
Depth Accuracy ~1mm Within 2 meters ~10mm Across All Distance
Display Resolution 1080P Below 360 x 240
Cost Medium High
Surrounding Environment Not Sensitive Very Sensitive To External Lighting

 

Check out our facial recognition case studies: Immigration Checkpoint at the iconic Hong Kong, Zhuhai, Macau Bridge.