3D Facial Model Generation from Non-Frontal Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing facial recognition systems are inefficient when comparing non-frontal face images, as they fail to accurately determine the resemblance between faces due to pose differences, leading to a significant loss in identification efficacy.
Innovation Solution
A method for generating a deformable three-dimensional face model from multiple images, involving the deformation of a facial template to minimize differences between characteristic points and surface features, while ensuring the deformed model corresponds to a human face, allowing for the estimation of pose and shape to generate a frontal view of the face.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a facial image is acquired from a non-frontal pose, then the acquisition system can capture the individual in natural viewing conditions, but the resemblance score between faces is substantially degraded
Solution Approach 1:
The patent transforms 2D facial images into a 3D space by constructing a three-dimensional model of the face. This dimensional transformation allows the system to account for pose variations and geometric distortions, enabling accurate resemblance comparison even when faces are captured from different angles. The 3D model includes depth information that compensates for the loss of information in 2D projections.
Solution Approach 2:
The patent changes the parameter space by representing facial features not just in 2D coordinates but in 3D spatial coordinates. By modeling the face as a three-dimensional object with x, y, z coordinates for each feature point, the system can transform between different pose representations and maintain measurement precision across varying acquisition conditions.
2Ease of manufacture
If statistical analysis with Gaussian density hypothesis is used to generate 3D face models, then the model generation process is simplified, but the model cannot accurately represent any human face due to the unproven hypothesis
Solution Approach 1:
Instead of generating 3D face models from statistical hypotheses, the patent creates accurate 3D representations by copying and mapping feature points from multiple 2D facial images. The system detects characteristic points in 2D images and uses triangulation and geometric constraints to reconstruct their 3D positions, thereby copying the actual facial geometry rather than approximating it through statistical distributions.
Solution Approach 2:
The patent replaces the statistical/mathematical hypothesis-based approach with a geometric/computational approach. Instead of assuming Gaussian density distributions, the system uses explicit geometric relationships, triangulation mathematics, and constraint satisfaction to compute 3D face models, substituting mechanical/computational methods for statistical assumptions.
Data Source
AI summary
The invention relates to a method for generating a three-dimensional facial model the shape of which can changed on the basis of a plurality of images of faces of persons, including the steps that involve: generating a facial template; acquiring shapes from examples of faces of persons; repeatedly changing the shape of the template for each example of a face of a person, so that the shape of the changed template corresponds to the shape of the face example, and determining the change in shape between the initial template and the changed template; and generating the facial model as a linear combination of the shape of the template and the changes in shape between the initial template and the changed template, for each example of a face of a person. The invention also relates to a method for processing an image of a face of a person such as to generate a three-dimensional image of the face of the person from of said deformable model.