Face Alignment via Local Feature Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face alignment methods using shape regression struggle with high training costs and generalization issues due to the large feature pool and noise in the feature pool, leading to suboptimal performance in testing.
Innovation Solution
The approach involves learning local features independently for each facial landmark and using the co-occurrence relationships among these features to refine the facial shape in a cascaded manner, applying a locality principle to focus on local texture information and shape context, rather than relying on the entire face region.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If shape regression uses a large feature pool from the entire face region, then the model can capture global facial structure information, but the training cost increases and the model suffers from noise and generalization issues
Solution Approach 1:
The patent divides the face region into multiple local patches around each facial landmark and processes them independently. Each patch is extracted and processed separately, then the results are combined to form the final shape regression output. This segmentation reduces the complexity of the overall feature pool while maintaining the ability to capture global structure through the combination of local features.
Solution Approach 2:
The patent applies different processing strategies to different local regions around each facial landmark. Instead of using a uniform feature pool for the entire face, it extracts and processes local features specifically around each landmark's expected position, tailored to the local structural characteristics. This local quality approach reduces noise and improves generalization by focusing on relevant local information.
2Productivity
If shape regression processes the entire face region as a whole, then global facial structure can be captured, but training time increases and the model becomes less efficient
Solution Approach 1:
The patent segments the face region into multiple local patches and processes them in parallel independently. This allows the model to train on smaller, more manageable feature sets for each landmark simultaneously, significantly reducing training time while maintaining accuracy through the combination of local results. The segmented approach enables efficient parallel processing and reduces the computational burden on training systems.
3Quantity of substance
If the model uses noise in the feature pool, then more information is available for training, but the noise degrades model performance in testing
Solution Approach 1:
The patent extracts and removes noisy features from the overall feature pool by processing only local regions around each facial landmark. The extraction process selectively keeps only the relevant local structural information while discarding irrelevant and noisy features from the entire face region. This results in a cleaner, more reliable feature set for training that generalizes better to testing data.
Solution Approach 2:
The patent applies local quality filtering by processing only the local patches around each landmark rather than the entire face. This local approach inherently filters out noise from regions that are not relevant to the current landmark, improving model robustness while still providing sufficient feature information for accurate shape regression.
Data Source
AI summary
The subject matter described herein relates to face alignment via shape regression. A method, computer storage medium, and system are provided. In one embodiment, the method comprises receiving an image including a face; and performing shape regression to estimate a facial shape in the image. For each stage in the shape regression, a local feature is extracted from a local region around each facial landmark in the image independently; and a joint projection is performed based on local features of multiple facial landmarks to predict a facial shape increment. Then, a facial shape of a current stage is generated based on the predicted facial shape increment and a facial shape of a previous stage.