Dual-Format Face Identification for Speed and Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face detection and recognition systems face challenges in balancing speed and accuracy, particularly in handling variations in facial expressions, lighting conditions, and accessory presence, leading to inefficiencies in processing and storage requirements.
Innovation Solution
The system employs a dual-similarity calculation approach using M-dimensional and N-dimensional feature vectors, where M-dimensional vectors prioritize recall ratio and N-dimensional vectors prioritize precision, reducing unnecessary calculations and storage needs by eliminating comparisons with non-matching vectors, and utilizing deep neural networks for feature extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single high-precision face detection algorithm is used, then detection accuracy is improved, but processing speed deteriorates
Solution Approach 1:
The face detection process is segmented into two distinct stages: a first face detection algorithm for initial detection and a second face detection algorithm for verification. This segmentation allows the system to use a fast algorithm for most cases while applying a more accurate algorithm only when needed, thus resolving the contradiction between speed and accuracy.
Solution Approach 2:
The system applies the computationally intensive second face detection algorithm only partially - specifically, only to detection results from the first algorithm that exceed a predetermined threshold. This partial application of the more accurate algorithm maintains high detection accuracy while avoiding the full computational cost of using it for all images, thereby improving processing speed.
2Measurement precision
If full feature vector comparison is performed for all registered vectors, then identification precision is improved, but computational load increases
Solution Approach 1:
The identification process is segmented into two phases: first comparing M-dimensional feature vectors to quickly eliminate non-matching candidates, then comparing N-dimensional feature vectors only for candidates that pass the first filter. This segmentation reduces computational load by avoiding full N-dimensional comparisons for all registered vectors while maintaining identification precision through the two-phase approach.
Solution Approach 2:
The system performs preliminary comparison using M-dimensional feature vectors before conducting the final N-dimensional comparison. This preliminary action filters out clearly non-matching candidates early in the process, reducing the number of vectors that require the computationally expensive N-dimensional comparison, thus lowering overall computational load while preserving identification precision.
3Measurement precision
If high-dimensional feature vectors are stored for all users, then identification accuracy is improved, but storage requirements increase
Solution Approach 1:
Feature vectors are segmented into two dimensional representations: M-dimensional vectors stored for all users and N-dimensional vectors stored only for users who pass the initial filtering stage. This segmentation allows the system to maintain high identification accuracy through N-dimensional vectors while minimizing storage requirements by using the more compact M-dimensional vectors for the broader user base.
Solution Approach 2:
The system stores the more space-consuming N-dimensional feature vectors only partially - specifically, only for users who are identified as potential matches during the M-dimensional comparison phase. This partial storage approach maintains identification accuracy for relevant users while significantly reducing overall storage requirements by not maintaining N-dimensional vectors for all users in the database.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An identification apparatus includes an extraction unit configured to extract an N-dimensional feature vector (N is an integer that is greater than M) in a second format including an M-dimensional feature vector (M is an integer of 2 or greater) in a first format from input data, and an identification unit configured to identify an identification target in the input data based on the feature vector in the first format and the feature vector in the second format.