Facial Feature Vector Matching with Multi-Task Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional facial recognition algorithms face challenges with high-dimensional facial feature vectors, leading to increased latency and reduced accuracy due to the curse of dimensionality, making it difficult to efficiently search and match vectors in large databases.
Innovation Solution
The implementation of a facial recognition system that combines hard and soft parameter sharing techniques using multi-task learning classifiers, reducing the search space by generating and utilizing smaller-sized facial feature vectors, typically between 100-300 dimensions, and performing vector matching with cosine similarity calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-dimensional facial feature vectors (512-4096 dimensions) are used for accurate facial recognition, then measurement precision is improved, but loss of time increases due to slower vector matching in large databases
Solution Approach 1:
The patent segments the facial recognition process into two stages: (1) a filtering stage using a lightweight classifier to quickly eliminate non-matching faces, and (2) a precise matching stage using cosine similarity on high-dimensional vectors for the remaining candidates. This segmentation allows the system to maintain high accuracy while reducing overall matching time by avoiding exhaustive comparison with all database vectors.
Solution Approach 2:
The patent performs preliminary classification using a trained model that predicts whether a query vector likely matches any database vector. This preliminary action filters out clearly non-matching cases before performing computationally expensive high-dimensional vector comparisons, thereby reducing latency while preserving accuracy for genuine matches.
2Ease of operation
If conventional vector matching algorithms (KD-trees, linear search) are used for high-dimensional vectors, then ease of operation is maintained, but productivity decreases due to inability to efficiently handle high dimensionality
Solution Approach 1:
The patent segments the vector matching task into a fast filtering operation using a linear classifier and a subsequent precise similarity computation. This segmentation enables the system to achieve high productivity by processing vectors through multiple stages with different computational characteristics, rather than relying on a single algorithm that must balance both speed and accuracy.
Solution Approach 2:
The patent changes the parameter representation by transforming facial images into fixed-dimensional vectors (e.g., 512 or 4096 dimensions) through neural network extraction, then applies cosine similarity computation. This parameter transformation enables efficient high-dimensional vector matching that overcomes the limitations of conventional algorithms like KD-trees which degrade in high dimensions.
3Measurement precision
If high-dimensional vector representations are used to capture detailed facial features, then measurement precision is improved, but device complexity increases due to computational requirements for handling and searching large vector spaces
Solution Approach 1:
The patent segments the computational workload by separating feature extraction (performed once during preprocessing) from similarity computation (performed during matching). This segmentation reduces device complexity during runtime by pre-computing and storing high-dimensional vectors, so that matching operations only require simple cosine similarity calculations rather than full neural network inference.
Solution Approach 2:
The patent performs preliminary vector extraction and normalization during an offline preprocessing stage, storing the resulting high-dimensional vectors in a database. This preliminary action shifts the computational complexity from the runtime matching system to the preprocessing stage, simplifying the device requirements for actual facial recognition operations.
Data Source
AI summary
A face recognition neural network generates small dimensional facial feature vectors for faces in received images. A multi-task learning classifier network receives generated vectors and outputs classifications for the corresponding images. The classifications are built by respective sub-networks each comprising classification layers arranged in levels. The layers in at least some levels receive as input intermediate outputs from an immediately upstream classification layer in the same sub-network and in each other sub-network. After training, a resultant facial feature vector corresponding to an input image is received from the neural network. Resultant classifiers for the input image are received from the classifier network. A database is searched for a subset of reference facial feature vectors having associated classification features matching the resultant classifiers. Vector matching identifies which one(s) of the reference facial feature vectors in the subset of reference facial feature vectors most closely match(es) the resultant facial feature vector.