Face Key Point Detection Using Segmented Neural Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing face key point detection methods using convolutional neural networks struggle with accuracy when dealing with face images that have large posture amplitudes or are partially blocked.

Innovation Solution

The proposed method improves face key point positioning accuracy by training a key point convolutional neural network model using face images with different posture information, and employing a network architecture that includes multiple target key point convolutional neural network models for enhanced feature extraction and alignment.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If a convolutional neural network is used for face key point detection, then the detection process can be automated and efficient, but the positioning accuracy deteriorates when dealing with face images that have large posture amplitudes or are blocked

Engineering Contradiction:
Improveautomation of face key point detectionVSAvoidface key point positioning accuracy
Core Design Contradiction:
Extent of automationVSMeasurement precision

Solution Approach 1:

The patent segments the face detection task into multiple specialized neural network models, each trained to handle specific posture conditions (e.g., frontal faces, profile faces, blocked faces). This segmentation allows each model to specialize in particular scenarios, thereby maintaining high automation while improving accuracy for diverse posture conditions that a single general model would struggle with.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system dynamically selects or switches between different neural network models based on the detected posture characteristics of the input face image. This dynamic adaptation enables the system to maintain high automation efficiency while achieving accurate key point positioning by using the most appropriate specialized model for each specific input condition.

Inventive Principle:
Principle #15Dynamics

2Adaptability or versatility

If training data includes face images with various posture information, then the model's adaptability to different postures improves, but the training complexity and data processing requirements increase

Engineering Contradiction:
Improvemodel adaptability to different posturesVSAvoidtraining data processing complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The training data is segmented into distinct categories based on posture characteristics (frontal, profile, blocked, etc.), and separate neural network models are trained for each category. This segmentation reduces the complexity of training a single model on all variations while improving adaptability, as each model focuses on learning specific posture patterns without interference from other categories.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary posture detection mechanism that first analyzes the input image to determine the posture category, then routes the image to the appropriate trained model. This intermediary layer simplifies the overall system complexity by providing a clear decision framework for selecting the right model, while enabling the system to handle diverse posture conditions with high adaptability.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3965003B1Image processing method and device
Publication Date: 2025.01.29 HUAWEI TECH CO LTD
  • EP3965003B1 patent drawingFigure 1~3A
  • EP3965003B1 patent drawingFigure 3B~4A
  • EP3965003B1 patent drawingFigure 4B

AI summary

An image processing method and apparatus are provided. The image processing method includes: obtaining a face image; separately obtaining a left face image and a right face image based on the face image; inputting the left face image into a first target key point convolutional neural network model, and outputting coordinates of a first left face key point, where the first target key point convolutional neural network model is obtained after a key point convolutional neural network model is trained by using a left face image having key point information; inputting the right face image into a second target key point convolutional neural network model, and outputting coordinates of a first right face key point, where the second target key point convolutional neural network model is obtained after a key point convolutional neural network model is trained by using a right face image having key point information; and obtaining coordinates of a face key point of the face image based on the coordinates of the first left face key point and the coordinates of the first right face key point. The image processing method may improve face key point positioning accuracy.