Face Image Normalization Using 3D Shape Model Posture Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing personal authentication techniques using face images face challenges in collating faces with different postures due to the need for special instruments for 3D shape information acquisition, high calculation costs, and susceptibility to erroneous feature point detection, which limits their use and accuracy.
Innovation Solution
An image processing device and method that estimates posture information including yaw and pitch angles from input images using a 3D shape model, allowing for the generation of normalized face images without requiring special instruments, by aligning feature points and correcting the orientation of faces to a standard posture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If three-dimensional shape information is acquired using special instruments such as a range finder, then the accuracy of face collation is improved, but the device complexity and ease of operation deteriorate due to the requirement of specialized equipment
Solution Approach 1:
The patent uses a pre-acquired three-dimensional shape model as a template to represent the target object's shape characteristics. Instead of requiring special instruments to measure the actual object, the system copies the essential geometric information from a standardized 3D model and applies it to normalize face images, thereby achieving accurate collation without specialized equipment
Solution Approach 2:
The three-dimensional shape model is prepared in advance and stored in the system. This preliminary preparation of geometric data allows the system to perform face normalization and collation without requiring real-time 3D scanning or special measurement instruments during the actual authentication process
2Measurement precision
If three-dimensional shape information is acquired and both two-dimensional image texture and three-dimensional shape information are processed, then the accuracy of face collation is improved, but the calculation cost increases
Solution Approach 1:
The patent separates the processing into distinct stages: first extracting feature points from the 2D image, then using the 3D shape model to determine normalization parameters, and finally applying geometric transformation. This segmentation allows the system to process only essential information at each stage rather than simultaneously processing all data, reducing overall computational burden
Solution Approach 2:
The system extracts only the necessary geometric parameters (such as yaw angle, pitch angle, and scale) from the three-dimensional shape model that are required for normalization. By taking out only the essential information needed for orientation correction rather than processing the complete 3D data set, the calculation cost is significantly reduced while maintaining collation accuracy
3Device complexity
If only four feature points (both eyes, nose, and mouth) are used for posture estimation, then the device complexity is reduced, but the reliability deteriorates due to high susceptibility to erroneous detection and outlier values
Solution Approach 1:
The patent applies different processing strategies to different parts of the face. Stable feature points (eyes, nose, mouth) are used for initial posture estimation, while additional feature points are utilized for verification and refinement. This local differentiation allows the system to leverage the stability of key landmarks while compensating for their vulnerability to errors through localized cross-validation
Solution Approach 2:
The system implements a feedback mechanism where the initially estimated posture parameters are used to predict the positions of other feature points, and these predicted positions are compared with actual detected positions. If discrepancies are found, the posture estimation is refined iteratively. This feedback loop significantly improves reliability by detecting and correcting erroneous detections without requiring a complete redesign of the feature point system
4Use of energy by moving object
If posture conversion is performed only on partial regions in the vicinity of feature points, then the calculation cost is reduced, but the manufacturing precision deteriorates due to great influence of erroneous feature point detection and limited global feature utilization
Solution Approach 1:
The patent transitions from two-dimensional image coordinate processing to three-dimensional shape model processing. By lifting the problem into 3D space, the system can perform global normalization that correctly handles perspective distortion and orientation changes across the entire face, not just local regions. The 3D shape model provides a comprehensive geometric framework that inherently accounts for global face structure, improving normalization accuracy without proportionally increasing calculation cost
Data Source
AI summary
An image processing device (10) includes a posture estimation unit (110) that estimates posture information including a yaw angle and a pitch angle of a person's face from an input image including the person's face, and an image conversion unit (120) that generates a normalized face image in which an orientation of a face is corrected, on the basis of positions of a plurality of feature points in a face region image which is a region including the person's face in the input image, positions of the plurality of feature points in a three-dimensional shape model of a person's face, and the posture information.