AI Facial Recognition With 3D Liveness Detection Against Spoofing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing facial recognition systems face challenges with scalability, cross-domain differences, and the limitations of 2D representations, leading to issues in liveness detection and recognition, especially with pose variations, facial expressions, and environmental effects, and are computationally expensive.
Innovation Solution
A biometric facial recognition system that utilizes both 2D and 3D features, including texture information and 3D face structures, combined with eye blink detection, to determine liveness and spoofing, using a pyramid-shaped neural network architecture like RetinaFace for alignment and an additive angular margin loss deep convolutional neural network (DCNN) for recognition, trained on live and spoof faces.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If holistic face-based detection is used to treat the whole-face regions or feature points, then feature extraction is simplified, but computational cost becomes very expensive
Solution Approach 1:
The patent segments the face detection process into multiple stages: initial coarse detection using simplified features, followed by progressive refinement in pyramid levels. This divides the computationally expensive holistic analysis into manageable segments that process only relevant regions at appropriate detail levels, reducing overall computational cost while maintaining detection accuracy.
Solution Approach 2:
The patent applies partial action by performing comprehensive holistic analysis only on detected face regions rather than entire images. The pyramid architecture applies full analysis only at necessary resolution levels, avoiding excessive computation on background or non-face areas, thus reducing computational cost while preserving detection effectiveness.
2Device complexity
If 2D facial recognition is used, then implementation is simpler, but pose variation, facial expression variation, lighting, and environmental effects remain unresolved
Solution Approach 1:
The patent transitions from 2D to 3D facial recognition by incorporating depth information through structured light projection and time-of-flight sensors. This dimensional enhancement enables the system to capture facial geometry that remains stable under pose variations, expression changes, and lighting conditions, significantly improving recognition reliability while maintaining manageable system complexity through established 3D sensing technologies.
3Adaptability or versatility
If more identities/faces are added to SoftMax classifier, then recognition coverage increases, but the number of parameters increases affecting performance
Solution Approach 1:
The patent employs 3D facial geometry and depth information as additional dimensions for discrimination, enabling the system to distinguish between more identities without proportionally increasing 2D parameter complexity. The depth maps and 3D point clouds provide unique geometric signatures that reduce parameter requirements while expanding recognition coverage.
Solution Approach 2:
The patent extracts and utilizes invariant 3D geometric features from faces, separating these stable structural elements from variable 2D appearance characteristics. By focusing classification on extracted 3D geometric descriptors rather than raw pixel data, the system achieves scalable recognition coverage with controlled parameter growth.
4Device complexity
If conventional 2D liveness detection is used, then implementation is straightforward, but spoofed images from internet or devices cannot be reliably distinguished
Solution Approach 1:
The patent uses 3D depth sensing to detect liveness by capturing the actual geometric structure of the face. Spoofed images from screens or photos lack the true 3D facial geometry and depth characteristics, allowing reliable distinction between live and spoofed subjects. This dimensional approach fundamentally improves liveness detection reliability beyond what 2D analysis can achieve.
Solution Approach 2:
The patent introduces structured light projection and time-of-flight sensing as intermediary mechanisms that actively illuminate the face and measure depth. These intermediaries provide additional physical measurements that reveal whether the subject is a live person or a spoof, enabling reliable liveness detection without substantially increasing system complexity.
Data Source
AI summary
A method, system, and computer readable medium are described to capture, detect, and recognize faces using machine learning and a single-stage face detector. A method to determine live faces from spoof 2D and spoof 3D images using a liveness score as well as a method to classify faces using machine learning deep convolutional neural networks is also described.