Human-Face Feature Extraction via Local Patch Fusion

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current video sequence-based human-face recognition methods face challenges due to frame blurring and low-resolution issues, leading to recognition errors, especially in dynamic environments, as they are not robust enough to handle pose variations and environmental influences.

Innovation Solution

A method that extracts human-face local feature vectors by dividing the face area into local patches, determining correspondence relationships between patches across frames, and combining these vectors to form a global feature vector, which is then used for recognition, employing techniques like optical flow and feature extraction methods such as LBP and PCA+LDA.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If image level fusion or super-resolution algorithms are used to enhance resolution, then recognition accuracy may improve, but computational complexity increases significantly

Engineering Contradiction:
Improverecognition accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the face image into multiple local patches (e.g., 8 patches) and processes each patch independently through feature extraction and fusion. This segmentation approach reduces the computational burden compared to processing the entire face image as a whole, while still maintaining recognition accuracy by capturing local facial characteristics.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent extracts features from only the necessary local patches rather than processing the entire face image. By selecting and processing only the relevant local regions (partial action), the system achieves sufficient recognition accuracy without the excessive computational complexity of full-image processing.

Inventive Principle:
Principle #16Partial or excessive action

2Productivity

If feature level fusion is used to extract local features, then processing efficiency improves, but the method is sensitive to pose variations and environmental influences

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidrobustness to pose variations
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent merges the feature vectors from multiple local patches through fusion operations (e.g., mean pooling, max pooling, or deep fusion) to create a comprehensive face representation. This merging process integrates information from different facial regions, making the recognition system more robust to pose variations and environmental factors while maintaining processing efficiency.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent employs different fusion parameters and weights for combining local feature vectors based on the specific characteristics of each patch and the overall face structure. By adjusting fusion parameters dynamically, the system optimizes its ability to handle pose variations and environmental influences while maintaining efficient processing.

Inventive Principle:
Principle #35Parameter changes

3Adaptability or versatility

If classifier level fusion is used for multi-scale recognition, then adaptability to different scales improves, but recognition accuracy decreases due to insufficient feature extraction

Engineering Contradiction:
Improvemulti-scale adaptabilityVSAvoidrecognition accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent extracts features at multiple scales by processing local patches of different sizes and resolutions. This multi-scale approach adds a dimensional aspect to the feature extraction process, enabling the system to adapt to different face sizes and poses while maintaining high recognition accuracy through comprehensive feature representation.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP2864933B1Method, apparatus and computer program product for human-face features extraction
Publication Date: 2019.09.25 NOKIA TECHNOLOGIES OY
  • EP2864933B1 patent drawingFigure 1
  • EP2864933B1 patent drawingFigure 2
  • EP2864933B1 patent drawingFigure 3

AI summary

The present invention provides a method for obtaining a human-face feature vector from a video image sequence, comprising: detecting a same human-face image in a plurality of image frames of the video sequence; dividing the detected human-face image into a plurality of local patches with a predetermined size, wherein each local patch is around or near a human-face feature point; determining a correspondence relationship between respective local patches of the same human-face image in the plurality of image frames of the video sequence; and using human-face local feature vector components extracted from respective local patches having a mutual correspondence relationship to form human-face local feature vectors representing facial points corresponding to the local patches. Besides, the present invention further provides an apparatus for obtaining a human-face feature vector from a video image sequence and a corresponding computer program product.