Convolutional Neural Network Feature Extraction for Face Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Face recognition technologies using principal component analysis (PCA) face challenges in recognition speed and reliability, especially under pose changes and facial expressions, and are not robust enough for illumination variations.
Innovation Solution
A deep learning method employing a neural network with multiple layers to extract low-level, mid-level, and high-level features from input images, allowing for robust feature recognition and verification by utilizing a convolutional neural network (CNN) structure to process and abstract image data effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If PCA is used to extract principal feature vectors from face images, then information dimension is reduced, but recognition speed and reliability are reduced
Solution Approach 1:
The patent segments the feature extraction process into multiple independent streams: low-level features (edges, textures) extracted by one neural network, mid-level features (facial components) by another, and high-level features (identity characteristics) by a third. This segmentation allows each stream to process different aspects of face recognition independently, avoiding the information loss inherent in PCA's single-stream dimensionality reduction while maintaining comprehensive feature representation for reliable recognition.
Solution Approach 2:
The patent transitions from PCA's single-dimensional feature space to a multi-dimensional feature hierarchy. Instead of projecting all face image information onto a single set of principal components, the system creates multiple feature dimensions at different levels of abstraction (low-level, mid-level, high-level), each captured by separate neural networks. This dimensional expansion preserves more information while enabling more reliable recognition through comprehensive feature integration.
2Device complexity
If PCA is used to extract principal feature vectors, then processing complexity is reduced, but recognition performance varies under pose changes and facial expressions
Solution Approach 1:
The patent implements dynamic feature extraction through multiple neural networks that adaptively capture different aspects of facial appearance. Rather than using static principal components from PCA, the system employs neural networks that can dynamically adjust to various poses and expressions by learning hierarchical features from training data. This dynamic approach allows the system to maintain recognition performance across diverse conditions while managing complexity through modular network architecture.
Solution Approach 2:
The patent changes the parameters of feature representation from PCA's fixed principal components to neural network parameters (weights and biases) that can be optimized for different recognition scenarios. By training neural networks on diverse pose and expression data, the system learns parameter transformations that are invariant to these variations, improving adaptability while the modular architecture keeps processing complexity manageable.
3Productivity
If feature information is reduced for faster processing, then recognition speed improves, but robustness against illumination variations deteriorates
Solution Approach 1:
The patent segments feature extraction into multiple specialized neural networks that process different feature types in parallel. Low-level feature networks capture illumination-invariant characteristics like edges and textures, while mid-level and high-level networks process semantic information. This segmentation allows the system to maintain recognition speed through efficient parallel processing while preserving robustness by dedicating specific networks to capture illumination-resistant features.
Solution Approach 2:
The patent replaces PCA's mechanical dimensionality reduction process with neural network-based feature learning. Instead of linearly projecting images to reduce dimensions, the system uses neural networks to learn non-linear feature transformations that are inherently more robust to illumination changes. The parallel architecture of multiple networks enables this substitution while maintaining processing speed through efficient computation.
Data Source
AI summary
A method of recognizing a feature of an image may include receiving an input image including an object; extracting first feature information using a first layer of a neural network, the first feature information indicating a first feature corresponding to the input image among a plurality of first features; extracting second feature information using a second layer of the neural network, the second feature information indicating a second feature among a plurality of second features, the indicated second feature corresponding to the first feature information; and recognizing an element corresponding to the object based on the first feature information and the second feature information.