Face Model Matrix Training via Grouped Covariance Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing face recognition technologies require terminals to load and process all face images in a library simultaneously, leading to high memory occupation and increased computational complexity.

Innovation Solution

Divide face images into groups, load and parse one group at a time, calculating intra-group and inter-group covariance matrices to train face model matrices, reducing memory usage and computational complexity.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If all face images in the face image library are loaded and processed simultaneously, then the face model matrix training accuracy is improved, but the memory occupation increases significantly

Engineering Contradiction:
Improveface model matrix training accuracyVSAvoidmemory occupation
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The face image library is divided into multiple groups, where each group contains a subset of face images. The system processes one group at a time, calculating intra-group and inter-group covariance matrices separately, then combines them to train the face model matrix. This segmentation approach maintains training accuracy while significantly reducing memory occupation, as only a portion of face images needs to be loaded into memory at any given time.

Inventive Principle:
Principle #1Segmentation

2Reliability

If all face images in the face image library are processed simultaneously, then the training completeness is improved, but the computational complexity increases

Engineering Contradiction:
Improvetraining completenessVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The training process is segmented into multiple passes, where each pass processes one group of face images. For each group, the system calculates the intra-group covariance matrix using only images in that group, then calculates the inter-group covariance matrix using mean vectors from all groups. This approach ensures training completeness while reducing computational complexity, as each processing step handles a manageable subset of data rather than the entire library at once.

Inventive Principle:
Principle #1Segmentation

3Quantity of substance

If face images are processed in groups, then the memory occupation is reduced, but the processing time may increase

Engineering Contradiction:
Improvememory occupationVSAvoidprocessing time
Core Design Contradiction:
Quantity of substanceVSLoss of time

Solution Approach 1:

The system processes face image groups in a continuous manner, where each group is processed immediately after the previous group. The intra-group and inter-group covariance matrices are calculated sequentially without idle time, and the face model matrix training continues uninterrupted through all groups. This continuous processing minimizes overhead and ensures that the reduced memory usage does not result in significant time loss.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP3382596B1Human face model matrix training method and apparatus, and storage medium
Publication Date: 2023.06.07 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3382596B1 patent drawingFigure 1
  • EP3382596B1 patent drawingFigure 2~3
  • EP3382596B1 patent drawingFigure 4~5

AI summary

A human face model matrix training method and apparatus. The method comprises: acquiring a human face image library, the human face image library comprising k groups of human face images, each group of human face images comprising at least one human face image of at least one person, k>2, and k being an integer (201); separately parsing each group of human face images among the k groups of human face images, and calculating a first matrix and a second matrix according to the parsing result, the first matrix being an intra-group covariance matrix of human face features of each group of human face images, and the second matrix being an inter-group covariance matrix of human face features of the k groups of human face images (202); and training a human face model according to the first matrix and the second matrix (203).