Face Comparison Device Correcting Occluded Feature Points
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face recognition techniques face challenges in accurately recognizing faces due to factors like angle, posture, and occlusion, leading to incorrect detection of feature points and requiring extensive learning data and manual input for correction.
Innovation Solution
A face comparison device that estimates occluded and pseudo-feature points using a three-dimensional stereoscopic face model, correcting feature points and generating normalized images for precise facial feature vector extraction and comparison.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If three-dimensional face model and geometric transformation parameters are used to generate face pattern images of different postures, then face recognition can be performed on disguised faces or old photographs, but feature points may be erroneously detected due to occlusion from angle and posture changes
Solution Approach 1:
The patent segments the feature point detection process into two distinct stages: first detecting feature points in the original input image, then detecting corresponding feature points in generated face pattern images. This segmentation allows the system to handle occlusion issues by comparing results from multiple viewpoints and selecting the most reliable feature point positions.
Solution Approach 2:
The patent transitions from two-dimensional face image analysis to three-dimensional face model processing. By generating face pattern images from a 3D model at different angles and postures, the system adds a dimensional aspect that enables viewing the face from multiple perspectives, thereby resolving occlusion problems that plague single 2D image analysis.
2Productivity
If feature points are detected from oblique face images, then face pattern images can be generated, but the detected feature points are located at incorrect positions due to self-occlusion
Solution Approach 1:
The patent implements a feedback mechanism where feature point detection results from the original image are used to guide and verify feature point detection in generated face pattern images. The system compares detected feature points across multiple images and uses this feedback to correct positional errors caused by occlusion, ensuring high accuracy in the final feature point positions.
Solution Approach 2:
The patent performs preliminary feature point detection on the original input image before generating face pattern images. This preliminary detection provides reference information that guides subsequent feature point detection in the generated images, preventing erroneous detections and ensuring accurate feature point positioning from the outset.
3Ease of operation
If geometric transformation parameters are calculated from detected feature points, then face images can be normalized, but the normalization is inaccurate when feature points are erroneously detected
Solution Approach 1:
The patent segments the normalization process into multiple steps: first calculating geometric transformation parameters from the original image feature points, then generating face pattern images with different postures, and finally performing normalization using feature points from both original and generated images. This segmentation allows for more accurate normalization by incorporating information from multiple sources.
Solution Approach 2:
The patent uses a composite approach by combining feature point information from the original image with feature point information from generated face pattern images. This composite use of multiple data sources creates a more robust normalization process that compensates for errors in individual detections, thereby improving overall normalization accuracy.
Data Source
AI summary
A face comparison device according to the present invention includes: an occluded-feature point estimation unit that estimates an occluded-feature point that is a feature point of an invisible face, and outputs position information of the occluded-feature point; a pseudo-feature point estimation unit that estimates a pseudo-feature point that is an erroneously detected feature point not captured within the input image due to occlusion, and outputs position information of the pseudo-feature point; and a comparison unit that generates a normalized image of the input image by using the position information of the occluded-feature point and the pseudo-feature point, generates a normalized image of an acquired comparison image, and compares the first facial feature vector extracted from the normalized image of the input image and the second facial feature vector extracted from the generated normalized image of the comparison image.