3D Face Model Generation Using Global and Local Feature Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating three-dimensional face models from two-dimensional images fail to accurately restore face details, as they primarily rely on global features without considering local features.
Innovation Solution
A method and apparatus that utilize a face model generation model trained on both global and local features of two-dimensional face images, integrating these features to generate a three-dimensional face model with improved detail restoration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If only global features are extracted from two-dimensional face images, then the generation process is simple and fast, but face details cannot be well restored
Solution Approach 1:
The patent divides the face image into multiple local regions (e.g., eyes, nose, mouth, cheeks) and extracts local features from each region separately using region-specific convolutional neural networks. This segmentation allows the system to capture detailed facial characteristics that would be lost in global feature extraction only, thereby improving face detail restoration quality while managing complexity through modular regional processing.
Solution Approach 2:
The patent applies different processing strategies to different facial regions by using multiple region-specific CNNs with different architectures and parameters tailored to each region's characteristics. For example, the eye region may use a deeper network to capture fine textures, while the mouth region uses a different configuration. This local quality approach ensures optimal feature extraction for each facial component, significantly enhancing overall face detail restoration.
2Measurement precision
If only global features are used to calculate three-dimensional face model parameters, then the calculation process is efficient, but the accuracy of face model generation is insufficient
Solution Approach 1:
The patent segments the 3D face model generation into multiple region-specific sub-tasks, where each region's 3D parameters are calculated independently based on its corresponding local features. This allows for more accurate region-specific parameter estimation while maintaining computational efficiency through parallel processing of multiple simpler regional models rather than one complex global model.
Solution Approach 2:
The patent combines the results from multiple region-specific 3D parameter calculations into a unified three-dimensional face model. By merging the locally-optimized regional models while maintaining their individual characteristics, the system achieves high overall accuracy that preserves both global face structure and local facial details, resolving the contradiction between model accuracy and system complexity.
3Manufacturing precision
If detailed local features are extracted from multiple convolutional layers, then face details are better captured, but the processing time and computational resources increase
Solution Approach 1:
The patent applies partial action by selectively extracting features from specific convolutional layers (e.g., early layers for basic structures, deeper layers for fine details) rather than processing all layers completely. This selective feature extraction captures essential face details while reducing the computational burden of processing every layer in full detail, thereby balancing accuracy with processing efficiency.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
A computer application method for generating a three-dimensional face model, comprising: obtaining a two-dimensional face image; invoking a face model generating model; inputting the two-dimensional face image into the face model generating model, and extracting a global feature and a local feature of the two-dimensional face image by means of the face model generating model; on the basis of the global feature and the local feature, obtaining a three-dimensional face model parameter; and on the basis of the three-dimensional face model parameter, outputting a three-dimensional face model corresponding to the two-dimensional face image.