Hierarchical Interlinked Multiscale Network for Face Parsing

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Image parsing, particularly for face images, is challenging due to poorly defined landmark points and the deformability of face parts, and existing methods struggle to effectively integrate multiscale information for accurate parsing.

Innovation Solution

A hierarchical interlinked multiscale convolutional neural network (HIM) framework is employed to generate and integrate image patches at different scales, enabling better utilization of data across various resolutions for locating and segmenting image components, including a level generator, interlinked combiner, and aggregator to refine component locations and footprints.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If landmark extraction is used to locate components, then component locations can be obtained, but the landmark points are not well-defined and uncertainty cannot be encoded

Engineering Contradiction:
Improvelandmark point definitionVSAvoiduncertainty encoding
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent segments the image into multiple scales (coarse to fine) and processes each scale separately through the cascaded hierarchical model. Instead of relying on single-scale landmark extraction, the system divides the parsing task across multiple resolution levels, where each level contributes to the final component localization and uncertainty estimation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent adds a scale dimension to the image parsing problem by processing images at multiple resolutions. The cascaded hierarchical model operates across this scale dimension, allowing the system to capture both global context (coarse scales) and local details (fine scales), thereby improving landmark definition and uncertainty encoding.

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

2Device complexity

If single-scale information is used for image parsing, then processing is simpler, but accurate parsing cannot be achieved due to inability to capture both global characteristics and local contours

Engineering Contradiction:
Improveprocessing complexityVSAvoidparsing accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent segments the image processing task across multiple scales, with each scale handling specific aspects of component detection. Coarse scales capture global characteristics while fine scales capture local contours, and the cascaded hierarchical model integrates these segmented scale-specific features to achieve accurate parsing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a scale dimension by processing images at multiple resolutions. This dimensional expansion allows the system to simultaneously capture global characteristics (at coarse scales) and local contours (at fine scales), resolving the trade-off between processing complexity and parsing accuracy.

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

3Ease of operation

If face parts are treated as rigid structures, then localization is easier, but accurate segmentation fails due to deformability of face parts

Engineering Contradiction:
Improvelocalization easeVSAvoidsegmentation accuracy
Core Design Contradiction:
Ease of operationVSMeasurement precision

Solution Approach 1:

The patent treats face parts as dynamic and deformable structures rather than rigid bodies. The cascaded hierarchical model at multiple scales can adapt to the deformable nature of face parts, capturing their varying shapes and configurations across different scales while maintaining accurate segmentation.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP3161728B1Hierarchical interlinked multi-scale convolutional network for image parsing
Publication Date: 2023.05.17 BEIJING KUANGSHI TECHNOLOGY CO LTD
  • EP3161728B1 patent drawingFigure 1
  • EP3161728B1 patent drawingFigure 2
  • EP3161728B1 patent drawingFigure 3

AI summary

A disclosed facial recognition system (and method) includes face parsing. In one approach, the face parsing is based on hierarchical interlinked multiscale convolutional neural network (HIM) to identify locations and/or footprints of components of a face image. The HIM generates multiple levels of image patches from different resolution images of the face image, where image patches for different levels have different resolutions. Moreover, the HIM integrates the image patches for different levels to generate interlinked image patches for different levels, where interlinked image patches for different levels have different resolutions. Furthermore, the HIM combines the interlinked image patches to identify refined locations and/or footprints of components.