DR-GAN Pose-Invariant Face Recognition via Disentangled Representation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current face recognition technologies face challenges in achieving accurate pose-invariant face recognition due to large pose discrepancies between face images, with existing methods either synthesizing frontal-pose images from non-frontal images or learning discriminative features from profile views, but these approaches are limited in handling unconstrained and real-world image sets effectively.
Innovation Solution
A Disentangled Representation Learning-Generative Adversarial Network (DR-GAN) framework is introduced, which uses a novel generator with an encoder-decoder structure to learn identity representations and synthesize faces at target poses, leveraging a pose code and random noise vector, while the discriminator predicts identity and pose, enabling the fusion of identity features from multiple images and handling image quality without direct supervision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning algorithms are used for face recognition, then recognition capability is improved, but pose-invariant performance deteriorates
Solution Approach 1:
The patent segments the face representation into two independent components: identity representation and pose representation. The identity representation captures facial characteristics invariant to pose, while the pose representation encodes orientation information. This segmentation allows the system to maintain high recognition accuracy while achieving pose invariance, resolving the contradiction between recognition capability and pose-invariant performance.
Solution Approach 2:
The patent introduces a pose code as an intermediary parameter that mediates between the input face image and the identity representation. The pose code explicitly models pose information separately from identity information, allowing the identity representation to remain pose-invariant while still accurately representing the facial identity. This intermediary approach enables both high recognition accuracy and pose invariance simultaneously.
2Measurement precision
If synthetic frontal-pose images are generated from non-frontal images, then frontal face recognition is improved, but generation accuracy deteriorates due to self-occlusion
Solution Approach 1:
The patent extracts the identity representation from the input face image while explicitly separating it from pose information. By taking out the identity components and representing them independently, the system can generate frontal face images that accurately preserve identity characteristics without being constrained by the self-occlusion problems of direct image synthesis methods. This extraction approach enables accurate frontal face generation from non-frontal images.
Solution Approach 2:
The patent changes the representation parameters from direct pixel-level image synthesis to a latent space representation model. Instead of directly generating frontal images from non-frontal images (which suffers from self-occlusion), the system transforms the input image into identity and pose parameters, then synthesizes the frontal image from these parameters. This parameter transformation approach resolves the self-occlusion limitation while maintaining generation accuracy.
3Adaptability or versatility
If multiple pose-specific models are trained and fused, then pose variation handling is improved, but model complexity increases
Solution Approach 1:
The patent creates a universal face representation model that simultaneously handles multiple pose variations through a single integrated architecture. Instead of training separate pose-specific models and fusing them (which increases complexity), the universal model learns to represent both identity and pose information in a unified manner, enabling it to handle various poses without requiring multiple specialized models. This multi-functionality approach maintains versatility while reducing model complexity.
Data Source
AI summary
A system and method for identifying a subject using imaging are provided. In some aspects, the method includes receiving an image depicting a subject to be identified, and applying a trained Disentangled Representation learning-Generative Adversarial Network (DR-GAN) to the image to generate an identity representation of the subject, wherein the DR-GAN comprises a discriminator and a generator having at least one of an encoder and a decoder. The method also includes identifying the subject using the identity representation, and generating a report indicative of the subject identified.