Face Recognition Robust to Alignment Using Deep Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional face recognition technologies are sensitive to face alignment variations, leading to misidentification of individuals due to errors in key point extraction and high computational complexity, especially when dealing with images of varying alignment states.
Innovation Solution
A deep neural network-based method that estimates prior facial shape information and feature information from input images, using a face image decoder to optimize alignment and extract facial shape-aware features, reducing the need for explicit face alignment during inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional face recognition pipelines perform explicit face alignment using key point extraction, then face images are normalized to a specific size, but face recognition performance deteriorates when there are errors in key point extraction or when images have various alignment states
Solution Approach 1:
The method performs preliminary face alignment by estimating facial shape prior information and using a face image decoder to generate an optimally aligned face image before feature extraction. This preliminary alignment action ensures that subsequent feature extraction operates on consistently aligned images, improving recognition reliability without requiring perfect key point extraction.
Solution Approach 2:
The patent introduces a face image decoder as an intermediary component between the input image and feature extraction. This decoder acts as a mediator that transforms variously aligned face images into a unified optimally aligned representation, thereby decoupling the feature extraction process from alignment variations and improving robustness.
2Reliability
If part-based face recognition is used to deal with alignment issues, then feature point information is used to select parts of the face, but the method still depends on feature point extraction performance and has alignment issues
Solution Approach 1:
The patent extracts and utilizes facial shape prior information separately from the main feature extraction process. By extracting this shape information independently and using it to guide alignment, the method reduces dependency on traditional feature point extraction while maintaining alignment robustness.
3Manufacturing precision
If studies increase the accuracy of feature point extraction, then face alignment improves, but computational complexity increases significantly
Solution Approach 1:
The patent replaces the traditional mechanical key point extraction system with a deep learning-based face image decoder. This substitution uses learned representations from training data to perform alignment, achieving high accuracy without the computational complexity of traditional feature point extraction algorithms.
4Ease of operation
If explicit face alignment processes are performed, then face images are normalized, but face recognition accuracy deteriorates when images have different alignment conditions due to key point extraction errors
Solution Approach 1:
The patent changes the parameters used for alignment by estimating facial shape prior information and using a decoder-based approach instead of traditional key point coordinates. This parameter change allows the system to handle various alignment conditions more reliably while maintaining ease of normalization.
Data Source
AI summary
A method and apparatus for face recognition robust to an alignment of the face comprising: estimating prior information of a facial shape from an input image cropped from an image including a face using the first deep neural network (DNN); extracting feature information of facial appearance from the input image by using a second DNN; training, by using a face image decoder based on the prior information and the feature information, the face recognition apparatus; and extracting, from a test image, facial shape-aware features in the inference step by using the trained second DNN.