Biometric Face Vector Quantization for Memory Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional biometric recognition methods face challenges due to the large size of reference biometric vectors, which occupy significant memory resources, and require large facial vectors for effective performance.
Innovation Solution
The method employs a neural network combined with quantification to create a biometric face vector, limiting the vector size while maintaining recognition performance, with a preferred embodiment using a neural network output layer of up to 128 neurons and a quantized biometric vector of less than 200 bytes, achieving a false rejection rate of 3% and false acceptance rate of 0.0001%.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional biometric recognition methods use large reference biometric vectors to achieve expected recognition performance, then recognition accuracy is improved, but memory resource consumption increases
Solution Approach 1:
The patent transforms the biometric vector from a large-dimensional continuous space to a compact quantized representation by changing the parameter dimensionality and data type. The reference biometric vector is quantized to 4-bit values and compressed to under 200 bytes, while the neural network output layer is configured with at most 128 neurons, fundamentally altering the parameter space to resolve the contradiction between accuracy and memory consumption
Solution Approach 2:
The patent extracts only the essential biometric features needed for recognition by using a neural network to generate a compressed representation. The quantification process extracts the most discriminative characteristics while discarding redundant information, enabling storage in limited memory resources while maintaining recognition performance
Data Source
AI summary
A biometric recognition method, comprising the step of calculating a similarity score of a candidate biometric vector with a reference biometric vector. At least one of the biometric vectors is extracted using at least one neural network.