Dual-Branch Neural Network Training for Low-Quality Face Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing face recognition technologies struggle with low-quality face images, as they fail to effectively extract and utilize the attention and gradient maps from such images, leading to degraded performance.

Innovation Solution

A dual-branch neural network model is employed, where one branch processes high-quality images and the other low-quality images, utilizing a softened attention-based gradient texture-guided learning method to align the attention and gradient maps, thereby improving recognition performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a neural network is trained to recognize high-quality face images, then recognition accuracy for high-quality images is improved, but recognition performance for low-quality images deteriorates

Engineering Contradiction:
Improverecognition accuracyVSAvoidperformance on low-quality images
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The neural network is divided into two separate models: a first model trained on high-quality images and a second model trained on low-quality images. Each model specializes in processing images of its respective quality level, allowing both to achieve optimal performance without interference. The segmentation enables the system to handle different image quality scenarios with dedicated architectures.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

An attention map serves as an intermediary between the first model and the second model. The attention map captures important features from high-quality images and guides the second model to focus on corresponding regions in low-quality images, effectively transferring useful information across quality boundaries without direct model coupling.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Productivity

If face features are extracted from low-quality images, then processing speed is maintained, but feature quality deteriorates due to damaged face features

Engineering Contradiction:
Improveprocessing speedVSAvoidfeature extraction quality
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The system performs preliminary processing by generating an attention map from a high-quality image before processing the low-quality image. This attention map pre-identifies important facial regions and guides the subsequent feature extraction process, ensuring that even with damaged features in low-quality images, the most critical information is captured first.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The gradient map acts as an intermediary that bridges the quality gap. It captures the structural information from the low-quality image while being guided by the attention map, allowing the system to maintain processing speed while improving feature extraction quality through guided gradient computation.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Ease of manufacture

If traditional training methods are used on low-quality images, then training simplicity is maintained, but recognition performance deteriorates

Engineering Contradiction:
Improvetraining simplicityVSAvoidrecognition performance
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The attention map serves as a guiding intermediary during training. It is generated from high-quality images and used to weight the loss function computation for the second model, directing the training process to focus on important facial regions. This approach maintains relative training simplicity while significantly improving recognition performance on low-quality images.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system introduces a weighted loss function that dynamically adjusts training parameters based on the attention map. By modifying the loss computation to emphasize important regions identified by the attention map, the training process adapts to prioritize learning from critical facial features, improving performance without fundamentally changing the training framework.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4202852B1Method and apparatus for training a neural network, image recognition method and storage medium
Publication Date: 2026.05.20 FUJITSU LTD
  • EP4202852B1 patent drawingFigure 1
  • EP4202852B1 patent drawingFigure 2
  • EP4202852B1 patent drawingFigure 3

AI summary

A method and an apparatus for training a neural network, an image recognition method and a computer readable storage medium are disclosed. The neural network includes a first model and a second model. The method for training a neural network includes: acquiring a second image from a first image, wherein a quality of the second image is lower than that of the first image; inputting the first image into the first model of the neural network, and inputting the second image into the second model of the neural network; calculating an attention map and a gradient map of the first model and an attention map and a gradient map of the second model; constructing a loss function based on a matrix of a dot product of the gradient map and the attention map of the first model and a matrix of a dot product of the gradient map and the attention map of the second model; and training the neural network by minimizing the loss function.