Iterative Training Set Selection for Facial Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Selecting an optimal set of people for training a classifier in facial recognition systems is challenging due to the need for diverse negative examples that resemble the user, particularly in terms of race, gender, and age, without available information on these attributes, and the computational burden of including all individuals in the training pool.
Innovation Solution
A system that selects an initial set of people for training, includes images of these individuals as both training and validation examples, and iteratively updates the training set based on false positive rates to improve classifier accuracy, using a combination of hardware and software processes to validate and retrain the classifier until stopping criteria are met.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all people in the pool are included in training, then the classifier accuracy is improved, but the computational cost becomes prohibitive
Solution Approach 1:
The training process is segmented into multiple iterative stages. In each iteration, only a selected subset of people (those with similar race, gender, and age to the user) is included in the training set, rather than training on all available people at once. This segmentation allows the system to progressively improve accuracy while keeping each training iteration computationally manageable.
Solution Approach 2:
The system applies partial action by selectively including only certain subsets of people (those demographically similar to the user) in each training iteration, rather than using all available training data. This partial inclusion is sufficient to improve accuracy for distinguishing similar-looking individuals without incurring the full computational cost of training on everyone.
2Reliability
If more people with similar race, gender, and age are included as negative examples, then the classifier's ability to distinguish similar-looking individuals is improved, but the complexity of selecting appropriate negative examples increases due to unavailable demographic information
Solution Approach 1:
The system performs preliminary classification in the validation stage to identify which people were incorrectly classified as the user. These misclassified individuals are then selected for inclusion in the next training iteration, eliminating the need for complex demographic analysis. The preliminary validation results directly guide the selection of appropriate negative examples.
Solution Approach 2:
The system uses feedback from the validation stage to iteratively improve the training set. By analyzing false positive classifications and using those results to select the next subset of people to train on, the system creates a closed-loop feedback mechanism that automatically identifies and addresses weaknesses in the classifier without requiring manual demographic analysis.
3Reliability
If the training set is updated iteratively based on false positives, then the classifier accuracy is improved, but the training time increases
Solution Approach 1:
The training process is divided into multiple smaller iterative stages, each focusing on a specific subset of challenging cases (people with similar demographics who were misclassified). By segmenting the training into targeted iterations rather than one large training pass, the system improves accuracy efficiently by concentrating computational resources on the most difficult distinguishing cases.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for selecting people for a training set. The methods, systems, and apparatus include actions of: obtaining an image of a user, obtaining an initial training set that includes the image of the user as a positive example and images of a subset of a set of other people as negative examples, training an initial classifier with the images of the initial training set, determining false positive classifications by the initial classifier, selecting people in the set of other people based on the false positive classifications, obtaining an updated training set that includes an image of the user, images of the subset of the set of other people, and images of the people that are selected, and generating an updated classifier with the updated training set.