Soft-Function Face Embedding Binarization for Fast Similarity Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The process of facial recognition using high-dimensional multibit floating number embeddings is time and resource-consuming due to the large number of reference embeddings, necessitating an efficient method for similarity comparison.
Innovation Solution
A method involving neural networks to generate compact binary representations of face recognition features using set bit location information, reducing complexity and memory usage by converting embeddings into binary form and utilizing efficient matching algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-dimensional multibit floating number embeddings are used for facial recognition, then recognition accuracy is improved, but computational complexity and processing time increase significantly
Solution Approach 1:
The patent transforms the embedding representation from high-dimensional multibit floating numbers to compact binary vectors with only 1 bit per dimension. This parameter transformation maintains the essential information needed for facial recognition while dramatically reducing computational complexity and memory requirements, directly resolving the contradiction between accuracy and complexity
Solution Approach 2:
The patent segments the embedding into binary dimensions where each dimension is independently binarized based on threshold comparison. This segmentation allows parallel processing and simplifies similarity computation through Hamming distance calculation, reducing the computational burden while preserving recognition capability
2Measurement precision
If high-dimensional multibit floating number embeddings are used for facial recognition, then recognition accuracy is improved, but memory requirements and processing time increase significantly
Solution Approach 1:
The patent changes the parameter representation from multibit floating numbers to single-bit binary values. This transformation reduces the data size by orders of magnitude, enabling faster memory access and significantly reducing processing time while maintaining the discriminative information for accurate facial recognition
Solution Approach 2:
The patent creates a simplified binary copy of the embedding that preserves the essential facial recognition information. This binary representation serves as an efficient proxy for the original high-dimensional embedding, enabling rapid similarity comparison through Hamming distance without requiring the full precision of the original data
3Measurement precision
If high-dimensional multibit floating number embeddings are used for facial recognition, then recognition accuracy is improved, but memory requirements increase significantly
Solution Approach 1:
The patent fundamentally changes the parameter storage requirement from multibit floating point numbers to single-bit binary values. This parameter compression reduces memory requirements by a factor of 8 or more per dimension while preserving the essential information needed for accurate facial recognition, directly addressing the memory constraint
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; applying a soft function on the multiple FR feature elements, by the one or more conversion layers, to provide a converted FR feature vector that comprises multiple converted FR feature elements that are non-binary; and generating a binary representation of the face recognition features based on the converted FR feature vector.