FACECOG – Face Recognition Solution for Heterogeneous IoT Platforms
Vicomtech’s FACECOG digital solution is a functionality from Vicomtech’s Viulib library, for recognizing person identities, based on their facial images, extracted from RGB image files, video files and video-cameras, or from multi-modal RGB and depth images obtained from depth cameras, such as those from Intel’s RealSense.
It can be used for video-surveillance applications such as registered people detection at a distance, and for user authentication. If multi-modal RGB and depth images are used, the solution provides additional mechanisms to avoid spoofing attacks.
The involved image processing stages are the following:
- Individuals are detected in captured images “in the wild”, extracting facial landmarks and regions, if visible;
- Facial image textures are normalized to minimize the impact of environmental and behavioural effects (illumination, image quality, distances from the camera, poses, expressions, etc);
- Representative, and privacy-preserving identity vectors (i-vectors) are extracted. These i-vectors are abstract representations of the person’s facial identity. The data is refined and updated as the system is used to improve the user’s experience, including when the user successfully authenticates.
No facial images are stored, neither during the user registration process, nor during the surveillance/authentication process. Only the extracted i-vectors are used, which do not leave the device where the solution has been deployed. The user’s facial images cannot be obtained from the extracted i-vectors, as these i-vectors are constructed by learning facial cues from facial image datasets built without any of the final users involved. Thus, the user’s privacy is totally preserved.