Hierarchical Patches Active Appearance Model for Face Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current face detection and alignment methods, such as Active Shape Models, Active Appearance Models, and Constrained Local Models, are not entirely satisfactory due to issues like sensitivity to local minima, occlusions, and limited adaptability to face variability, particularly in handling texture convergence and mesh deformation.

Innovation Solution

The proposed method employs Hierarchical Patches Active Appearance Models (HPAAM) that converge the shape and texture of a hierarchy of statistical models on real data, using a learning phase with principal component analysis and a detection phase with a compositional inverse algorithm to optimize shape, pose, and texture parameters, effectively addressing the limitations of existing methods.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If Active Shape Models (ASM) are used for face detection, then the detection process is simple and fast, but the method is sensitive to local minima and occlusions

Engineering Contradiction:
Improvedetection speedVSAvoidrobustness to occlusions and local minima
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The face detection process is segmented into multiple hierarchical levels, starting with detecting prominent facial features (eyes, nose, mouth) and progressively refining to detect less prominent features. This hierarchical segmentation allows the system to build up detection confidence gradually, reducing sensitivity to local minima and occlusions while maintaining computational efficiency.

Inventive Principle:
Principle #1Segmentation

2Reliability

If Active Appearance Models (AAM) are used for face detection, then robustness to local minima is improved, but texture convergence is problematic due to mesh deformation issues

Engineering Contradiction:
Improverobustness to local minimaVSAvoidmesh deformation complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The appearance model is segmented into shape components and texture components that are optimized separately. The shape model handles geometric variations using hierarchical feature detection, while the texture model handles appearance variations independently. This segmentation eliminates the need for complex mesh deformation to achieve texture convergence, as textures are mapped based on detected feature positions rather than continuous mesh warping.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The problem is transformed from a continuous 2D mesh deformation problem into a discrete feature-point based approach. Instead of deforming a continuous texture map across the entire face, the system uses independently detected feature points to sample and match texture regions, effectively moving from a spatially continuous problem to a set of discrete correspondence problems that are computationally simpler.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Reliability

If Constrained Local Models (CLM) are used for face detection, then robustness to occlusions is improved, but the method is more sensitive to local minima

Engineering Contradiction:
Improverobustness to occlusionsVSAvoidsensitivity to local minima
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The face is segmented into multiple independent feature regions (eyes, nose, mouth) that are detected hierarchically. Each region is detected using local image gradients and statistical models, allowing occlusions in one region to be compensated by successful detection in other regions. This segmentation approach provides robustness to occlusions while the hierarchical refinement process reduces sensitivity to local minima by progressively improving detection accuracy.

Inventive Principle:
Principle #1Segmentation

4Adaptability or versatility

If statistical models converge shape and texture jointly using gradient descent, then adaptability to facial variability is improved, but the mesh deformation function must always create a smooth mesh which is problematic

Engineering Contradiction:
Improveadaptability to facial variabilityVSAvoidmesh deformation smoothness constraint
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The joint shape and texture optimization is segmented into separate detection stages. Shape parameters are optimized first using gradient-based methods on shape gradients, followed by texture parameter optimization using the detected shape as a constraint. This segmentation removes the need for continuous smooth mesh deformation during optimization, as shape and texture are fitted in sequence rather than simultaneously, maintaining adaptability while eliminating mesh smoothness constraints.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP2678804B1Method for detecting a predefined set of characteristic points of a face
Publication Date: 2024.04.10 FITTINGBOX
  • EP2678804B1 patent drawingFigure 1~2

AI summary

The invention relates to a method for detecting a predefined set of characteristic points of a face from an image of said face. The method comprises a step of causing the shape and/or the texture of a hierarchy of statistical models of portions of a face to converge on the real data supplied by the image of the face.