Hybrid Video Face Recognition System for Embedded Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face recognition systems on mobile devices face challenges in achieving high accuracy while maintaining fast response times and low power consumption, especially due to pose and illumination variations, which are not effectively addressed by current deep-learning based systems given the limited processing capabilities of modern mobile processors.
Innovation Solution
A novel face recognition system combines deep-learning approaches with conventional machine learning classifiers, utilizing a robust convolutional neural network (CNN) for feature extraction and a specialized training procedure to ensure Gaussianity of mathematical face descriptors, along with a compressed CNN architecture that reduces computational complexity and power consumption, and employs sequential images to address pose and illumination variability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep-learning based face recognition systems are used to achieve high recognition accuracy, then recognition accuracy is improved, but processing speed and power consumption worsen due to computational complexity
Solution Approach 1:
The face recognition system is segmented into two distinct modules: a deep-learning-based feature extraction module (using CNNs) and a conventional machine learning classification module. This segmentation allows the computationally intensive feature extraction to be performed once, while the lighter classification process can be executed rapidly for real-time recognition, thus resolving the contradiction between accuracy and speed.
Solution Approach 2:
The system performs preliminary action by pre-training the CNN to extract robust facial features and pre-computing the classification models offline. This preliminary processing creates optimized feature representations that can be quickly classified during runtime, achieving high accuracy without compromising processing speed during actual face recognition operations.
2Measurement precision
If deep-learning based face recognition systems are used to achieve high recognition accuracy, then recognition accuracy is improved, but power consumption worsens due to computational complexity
Solution Approach 1:
The system segments the computational workload into a heavy offline training phase (feature extraction with CNN) and a light online recognition phase (classification with conventional ML). This segmentation ensures that high power consumption is confined to initial setup, while runtime operation consumes minimal power, resolving the contradiction between accuracy and power consumption.
Solution Approach 2:
The system uses lightweight conventional machine learning classifiers instead of continuous deep-learning inference, effectively replacing expensive computationally intensive operations with cheaper, simpler operations during runtime. This allows the system to achieve high accuracy through the pre-trained CNN while maintaining low power consumption during actual recognition tasks.
3Productivity
If conventional machine learning classifiers are used to achieve fast processing speed and low power consumption, then processing speed and power efficiency are improved, but recognition accuracy worsens
Solution Approach 1:
The system merges the strengths of both deep-learning-based feature extraction and conventional machine learning classification. The CNN provides robust, accurate feature representations while the conventional classifier delivers fast, efficient decision-making. This combination achieves both high recognition accuracy and fast processing speed simultaneously.
Solution Approach 2:
The CNN acts as an intermediary that transforms raw facial images into optimized feature vectors. These intermediate feature representations are specifically designed to be highly discriminative yet compact, enabling conventional classifiers to achieve high accuracy with minimal computational effort, thus resolving the contradiction between speed and accuracy.
4Speed
If face recognition systems process single images to achieve fast response time, then response speed is improved, but recognition accuracy worsens due to pose and illumination variations
Solution Approach 1:
The system performs preliminary action by processing multiple sequential frames and accumulating evidence before making a recognition decision. This preliminary processing of multiple frames compensates for pose and illumination variations, improving accuracy without significantly increasing response time, as the accumulation process is optimized for speed.
Solution Approach 2:
The system maintains continuous processing of facial features across multiple frames, continuously updating the recognition confidence. This continuous action allows the system to leverage temporal information and maintain high accuracy even with single-frame limitations, while still achieving fast response times through efficient frame-by-frame processing.
Data Source
AI summary
An exemplary embodiment relates to the field of Automatic Face Recognition (AFR) systems. More specifically one exemplary embodiment relates at least to a method and a system capable of recognizing the face of a person using a device equipped with a camera of any kind and an associated computer, such as an embedded computer. The system is alternatively suitable to be implemented as an embedded system with minimal processing hardware capabilities, consuming very low power.