Cascade Face Key Point Detection for Mobile Resource Constraints

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing biometric processing technologies face challenges in accurately and efficiently detecting face key points on mobile devices with scarce resources, as larger network models and hierarchical models require significant computation and resources, making real-time detection difficult.

Innovation Solution

A method involving a cascade network structure with three-level processing, including a face key point detecting model, a posture angle estimating model, and a position correcting model, which distinguishes face types to simplify positioning and reduce computational load, ensuring accurate and real-time detection of face key points.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a larger network model is introduced to improve detection accuracy, then the accuracy in detecting face key points is improved, but the computational load and resource consumption increase, making real-time detection difficult on mobile devices

Engineering Contradiction:
Improveaccuracy in detecting face key pointsVSAvoidcomputational load and resource consumption
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent divides the face detection task into multiple stages using a cascade network structure. The first stage uses a simpler model for rough detection, and subsequent stages progressively refine the results with more complex models. This segmentation allows the system to achieve high accuracy while keeping the computational load of each individual stage manageable for mobile devices.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent employs adaptive model selection based on detected face types. Different face types (e.g., frontal, profile, angled) are identified and matched with appropriately complex detection models. This dynamic adaptation ensures that the system uses only the necessary computational resources for each specific case, rather than always deploying the most complex model.

Inventive Principle:
Principle #15Dynamics

2Measurement precision

If a more hierarchical network model is introduced to improve detection accuracy, then the accuracy in detecting face key points is improved, but the time complexity increases, making real-time detection difficult

Engineering Contradiction:
Improveaccuracy in detecting face key pointsVSAvoidtime complexity
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The cascade network structure segments the detection process into multiple passes, where each pass processes only the regions of interest identified in previous passes. This reduces the overall time complexity compared to applying a single complex hierarchical model to the entire image, while still achieving high accuracy through progressive refinement.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary face type classification and rough key point detection before applying more complex correction models. This preliminary action narrows down the search space and reduces the computational time required for subsequent detailed detection, enabling real-time performance.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If a larger network model is introduced to improve detection accuracy, then the accuracy in detecting face key points is improved, but the demand on resources increases, making it difficult to apply to mobile terminals

Engineering Contradiction:
Improveaccuracy in detecting face key pointsVSAvoiddemand on resources
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

By segmenting the detection task into multiple stages with progressively complex models, the patent reduces the resource demand of each individual model stage. The first stage uses a lightweight model that consumes minimal resources, while subsequent stages process only the refined regions, thereby distributing resource consumption across multiple lighter components rather than requiring one large resource-intensive model.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies full computational resources only when necessary. For example, complex correction models are applied selectively based on face type and detection confidence levels. In many cases, simpler models provide sufficient accuracy, allowing the system to conserve resources while maintaining high accuracy when needed.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3882809B1Face key point detection method, apparatus, computer device and storage medium
Publication Date: 2024.10.16 BIGO TECH PTE LTD
  • EP3882809B1 patent drawingFigure 1
  • EP3882809B1 patent drawingFigure 2
  • EP3882809B1 patent drawingFigure 3

AI summary

Provided are a face key point detection method, apparatus, computer device and storage medium, wherein the method comprises: obtaining image data containing a face; detecting a first face key point of the face in the image data; identifying a face type of the face according to the first face key point; performing position correction on the first face key point with reference to the face type; extracting facial features image containing facial features from the image data based on the first face key point after position correction; and detecting a second face key point of the face in the facial features image.