Object Pose Normalization via Smoothing and Affine Synthesis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for normalizing object poses, particularly in facial images, face challenges with computational complexity, self-occlusion, and non-rigid deformation, limiting the effectiveness of face recognition systems.
Innovation Solution
An object pose normalization method and apparatus that performs smoothing transformation on non-frontal images to generate a smoothed object image, using pose determination and thin plate spline algorithms to synthesize a frontal image, addressing issues of non-rigid deformation and occlusion while maintaining computational efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If 3D morphable model is used to synthesize facial images, then pose normalization can be achieved, but computational complexity increases due to optimization of large number of parameters
Solution Approach 1:
The patent extracts and focuses only on the essential 2D affine transformation parameters (rotation, translation, scaling) rather than optimizing all 3D morphable model parameters. This selective extraction reduces computational complexity while maintaining sufficient pose normalization precision for facial recognition applications.
Solution Approach 2:
Instead of using complex 3D models to achieve pose normalization, the patent inverts the approach by using simpler 2D affine transformations on normalized 2D images. This inversion of the traditional 3D-based approach achieves the same normalization goal with significantly reduced computational complexity.
2Device complexity
If 2D approaches with Affine transformation are used, then computational complexity is reduced, but non-rigid deformation and self occlusion cannot be compensated
Solution Approach 1:
The patent applies different processing strategies to different regions of the facial image. Local feature point detection and matching are performed on specific landmark regions, while global affine transformation is applied to the overall image structure. This localized approach enables compensation of non-rigid deformations in critical facial regions while maintaining computational efficiency.
Solution Approach 2:
The patent introduces an intermediary step of detecting and matching local feature points (such as eyes, nose, mouth landmarks) between source and target images. These feature points serve as intermediaries that guide the affine transformation, enabling the system to handle non-rigid deformations and occlusions that would otherwise be impossible to compensate for with simple 2D approaches.
3Measurement precision
If feature points are detected using AAMs or ASMs, then facial feature localization is achieved, but initialization and automatic localization become difficult
Solution Approach 1:
The patent performs preliminary detection of facial feature points using robust methods (AAMs or ASMs) on the source image before transformation. These pre-detected feature points serve as anchors that guide subsequent feature point detection in the target image, making automatic localization more reliable and easier to implement without requiring complex initialization procedures.
Data Source
AI summary
An object pose normalization method and apparatus and an object recognition method are provided. The object pose normalization method includes: determining a pose of a non-frontal image of an object; performing smoothing transformation on the non-frontal image of the object, thereby generating a smoothed object image; and synthesizing a frontal image of the object by using the pose determination result and the smoothed object image. According to the method and apparatus, a front object image can be synthesized by using a non-frontal object image without causing an image distortion problem due to self-occlusion and non-rigid deformation.