Dense Connection Face Recognition Network with Bottleneck Modules
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current face identification technologies face challenges with low identification rates due to feature loss during layer-by-layer extraction and high dimensionality of feature maps in deep Resnet networks, which also lead to slow computing speeds and memory issues, making them unsuitable for embedding in terminals.
Innovation Solution
A neural network model incorporating a dense connection policy with four bottleneck networks connected in series, using 1*1 convolutional layers to reduce data dimensions and improve computing speed, and employing feature fusion to enhance face feature extraction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a residual network (Resnet) is constructed to carry out feature fusion, then feature extraction capability is improved, but the feature map has too many dimensions, leading to low computing speed and high memory requirements
Solution Approach 1:
The network is divided into multiple stages, each containing bottleneck modules with specific functions. The segmentation allows progressive feature extraction while controlling dimensionality at each stage, preventing the feature map from becoming too high-dimensional overall.
Solution Approach 2:
The bottleneck modules use 1x1 convolutions to change the dimensionality of feature maps by projecting them to lower-dimensional spaces. This dimensionality reduction is performed strategically at key points in the network to maintain computing efficiency while preserving essential features.
2Measurement precision
If a residual network (Resnet) is constructed to carry out feature fusion, then feature extraction capability is improved, but memory requirements increase making it impossible to embed in terminal devices
Solution Approach 1:
The network architecture is segmented into modular stages with bottleneck modules that can be selectively configured. This segmentation allows the model to achieve high feature extraction capability while keeping the total parameter count and memory footprint suitable for terminal device embedding.
Solution Approach 2:
The bottleneck modules employ parameter sharing and dimensionality reduction techniques to maintain high feature extraction capability with fewer parameters. The 1x1 convolution layers strategically reduce the number of parameters needed while preserving essential feature representations.
3Measurement precision
If face features are extracted layer by layer through a trained neural network model, then a final feature map is obtained, but a part of face features are lost, resulting in low identification rate
Solution Approach 1:
The dense connection structure merges feature maps from multiple previous layers through concatenation, allowing the network to retain and utilize features from all stages of processing. This merging prevents feature loss by ensuring that early-layer features are preserved and combined with later-layer abstract features.
Solution Approach 2:
The dense connection architecture provides feedback pathways where feature maps from earlier layers are fed forward to subsequent layers. This feedback mechanism ensures that important low-level features are not lost during the progressive extraction process and can be combined with high-level semantic features.
4Productivity
If the feature map dimensionality is reduced using 1*1 convolutional layers, then computing speed is improved, but the network structure becomes more complex
Solution Approach 1:
The bottleneck modules are designed as universal building blocks that can be applied repeatedly across different stages of the network. Each module integrates dimensionality reduction, feature extraction, and transition functions, reducing the need for separate components and simplifying the overall structural complexity despite the enhanced functionality.
Data Source
AI summary
Embodiments of the present application disclose a deep learning-based face identification method and a related product. The method is applied to an electronic apparatus. The method includes: acquiring an aligned face image, and scaling the face image at a preset ratio to obtain a target image; extracting a pixel matrix of the target image, inputting the pixel matrix to the neural network model, and executing multilayer computing to obtain a computing result; and calculating a cosine similarity degree between the computing result and a facial template vector, wherein comparison is successful if the cosine similarity degree is greater than a preset threshold. The embodiments of the present application help improve the identification success rate and the identification speed.