Similarity-Embedding Binarization for Faster Face Recognition Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The process of facial recognition using high-dimensional FR embeddings is time and resource-consuming due to the calculation of similarity between large numbers of reference embeddings, which is inefficient and burdensome on hardware resources.
Innovation Solution
A method involving a neural network that generates a compact binary representation of face recognition features using set bit location information, reducing complexity and memory requirements by treating embeddings as a collection of active features, and employing efficient matching algorithms from textual search-engine domains.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-dimensional multibit floating point FR embeddings are used for facial recognition, then accuracy is improved, but computational complexity and memory usage increase significantly
Solution Approach 1:
The patent changes the parameter representation from high-dimensional multibit floating point numbers to compact binary vectors. Each embedding element is converted from a continuous floating-point value to a discrete binary value (0 or 1), fundamentally altering the data representation parameters. This parameter transformation maintains the essential discriminative information needed for accurate facial recognition while dramatically reducing computational complexity and memory requirements.
Solution Approach 2:
The patent extracts only the essential discriminative features from the full high-dimensional floating-point embeddings. By converting to binary representation, the system extracts and retains only the most significant feature bits that contribute to facial recognition accuracy, discarding redundant precision information. This extraction process maintains measurement precision while reducing device complexity.
2Measurement precision
If high-dimensional multibit floating point FR embeddings are used for facial recognition, then accuracy is improved, but memory usage increases significantly
Solution Approach 1:
The patent fundamentally changes the memory storage parameters by converting from multibit floating-point representation to binary representation. Each embedding element that previously required multiple bytes for floating-point storage is reduced to a single bit in binary form. This parameter change achieves dramatic memory compression while preserving the essential information needed for accurate facial recognition through the binary feature representation.
Solution Approach 2:
The patent extracts and retains only the essential discriminative information from the full-precision floating-point embeddings. By representing features in binary form, the system extracts the most significant bits that contribute to recognition accuracy while eliminating redundant precision data. This extraction achieves memory efficiency without sacrificing measurement precision.
3Adaptability or versatility
If similarity calculation is performed between large numbers of reference embeddings, then recognition capability is improved, but processing time increases significantly
Solution Approach 1:
The patent replaces the traditional mechanical computation of similarity calculations (inner products of high-dimensional floating-point vectors) with a simplified binary comparison mechanism. Instead of performing complex floating-point arithmetic operations, the system uses binary vector comparisons that can be executed much faster using bitwise operations. This substitution maintains recognition capability by preserving the essential feature relationships while dramatically reducing processing time through more efficient computational mechanics.
Solution Approach 2:
The patent changes the computational parameters from floating-point arithmetic operations to binary comparison operations. By representing embeddings in binary form, the similarity calculation transforms from complex mathematical operations to simpler bitwise operations that are inherently faster to execute. This parameter change maintains the ability to distinguish between different faces while reducing processing time through more efficient computational parameters.
Data Source
AI summary
Computerized systems, and method and computer readable media. The method may include receiving, by a neural network, input face visual information; wherein the neural network comprises multiple convolutional layers, an embedding layer and one or more conversion layers; generating, by the embedding layer, a face recognition (FR) feature vector that comprises multiple FR feature elements; and generating a binary representation of the face recognition features based on the FR feature vector.