Dual-Branch Neural Network Training for Low-Quality Face Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face recognition technologies struggle with low-quality images, as damaged face features degrade the quality of extracted features, leading to poor recognition performance.
Innovation Solution
A dual-branch neural network model is trained using a softened attention-based gradient texture-guided learning method, where a high-quality image is input into one branch and a low-quality image derived from it is input into another, with a loss function constructed based on the dot product matrices of gradient and attention maps to minimize differences and enhance feature consistency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a standard face recognition model is used, then the recognition performance on high-quality images is good, but the recognition performance on low-quality images degrades due to damaged face features
Solution Approach 1:
The neural network is divided into two separate branches: a first branch that processes high-quality images and a second branch that processes low-quality images. Each branch independently extracts features from images of different qualities, allowing the system to handle degraded features in low-quality images separately without being constrained by the requirements for high-quality image processing.
Solution Approach 2:
A loss function is introduced as an intermediary mechanism that guides the training process by measuring the difference between features extracted from high-quality and low-quality images. This loss function acts as a mediator that enables the model to learn how to compensate for feature degradation in low-quality images by comparing it with the corresponding high-quality features.
2Productivity
If features are extracted from low-quality images, then processing can proceed, but the extracted feature quality degrades leading to poor recognition
Solution Approach 1:
The system segments the feature extraction process into two parallel pathways: one for high-quality images and one for low-quality images. This segmentation allows each pathway to be optimized for its specific input quality, enabling continuous processing of low-quality images while maintaining feature extraction quality through separate optimization.
Solution Approach 2:
The model parameters are optimized separately for each branch through the proposed loss function. By changing the training parameters and objective function to account for the specific characteristics of low-quality images, the system achieves better feature extraction quality from degraded inputs without compromising the processing capability.
3Adaptability or versatility
If a single model processes both high and low quality images, then device complexity is low, but recognition performance on low-quality images is limited
Solution Approach 1:
The neural network is segmented into two parallel branches that process high-quality and low-quality images separately. This segmentation increases adaptability by allowing each branch to specialize in handling images of its respective quality level, while the overall structure remains relatively simple and manageable.
Solution Approach 2:
The dual-branch neural network structure serves multiple functions: it can process both high-quality and low-quality images simultaneously, and it can be trained using a unified loss function that optimizes performance across both image qualities. This multi-functionality achieves broad adaptability without requiring completely separate systems.
Data Source
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.