Lightweight Facial Alignment Network via One-Shot NAS
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing facial alignment networks (FANs) achieve significant accuracy but come with larger model sizes and expensive computation costs, making them infeasible for real-time applications on edge devices such as smartphones and tablets.
Innovation Solution
The implementation of a model compression approach using One-Shot Neural Architecture Search (NAS) for FANs, which involves Expand & Shrink Training (EST) operations to generate candidate submodels and perform evolutionary searches for optimal model selection, resulting in a lightweight and efficient FAN model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If facial alignment networks are designed to achieve high accuracy on public datasets, then accuracy is improved, but model size increases and computation costs become expensive
Solution Approach 1:
The patent divides the large facial alignment network into multiple smaller candidate submodels with different architecture configurations. These submodels are segmented by varying parameters such as number of layers, filter sizes, and network depth, allowing the system to select an appropriate subset for deployment on edge devices with limited resources.
Solution Approach 2:
The patent systematically changes architectural parameters of the neural network including depth, width, kernel sizes, and stride values to generate a family of candidate models. By exploring different parameter combinations, the system identifies models that achieve high accuracy while maintaining small size suitable for edge deployment.
2Measurement precision
If facial alignment networks are designed to achieve high accuracy on public datasets, then accuracy is improved, but computation costs become expensive
Solution Approach 1:
The patent employs dynamic model selection where the system can adaptively choose different candidate submodels based on available computational resources and performance requirements. The evolutionary search algorithm dynamically evaluates candidate models and selects the optimal configuration that balances accuracy with computational efficiency for real-time edge processing.
Solution Approach 2:
The patent optimizes computational parameters including reducing the number of convolutional layers, adjusting filter counts, and modifying kernel sizes to decrease FLOPs and memory usage. These parameter changes enable the model to run efficiently on mobile devices with limited computational power while maintaining acceptable accuracy levels.
3Speed
If manually designed mobile-targeted models are used to reduce computational time, then speed is improved, but accuracy becomes insufficient compared to carefully designed face alignment models
Solution Approach 1:
The patent creates a universal model family that can serve multiple functions - achieving both high speed and high accuracy depending on the specific application requirements. The evolutionary search generates candidate models optimized for different trade-offs, allowing the system to select models that simultaneously satisfy speed constraints and accuracy requirements, unlike manually designed models that must prioritize one aspect.
Solution Approach 2:
The patent uses evolutionary search with feedback mechanisms that evaluate candidate models based on both accuracy metrics and computational performance. The feedback from evaluation results guides the selection and refinement of candidate submodels, ensuring that the final selected model achieves the desired balance between speed and accuracy for real-time face alignment on edge devices.
4Measurement precision
If large facial alignment networks are deployed on edge devices, then accuracy is maintained, but real-time processing capability is lost
Solution Approach 1:
The patent segments the large network into multiple candidate submodels with progressively reduced complexity. This segmentation allows the system to deploy lightweight models on edge devices that can process images in real-time, while still maintaining high accuracy through the selection of optimized architectures from the candidate set.
Solution Approach 2:
The patent implements dynamic model adaptation where the system can switch between different candidate submodels based on real-time performance requirements. The evolutionary search identifies models with optimized inference speed, enabling real-time processing on mobile devices while maintaining accuracy through the use of carefully selected architectural configurations.
Data Source
AI summary
With Convolutional Neural Networks (CNN), facial alignment networks (FAN) have achieved significant accuracy on a wide range of public datasets, which comes along with larger model size and expensive computation costs, making it infeasible to adapt them to real-time applications on edge devices. There is provided a model compression approach on FAN using One-Shot Neural Architecture Search to overcome this problem while preserving performance criteria. Methods and devices provide efficient training and searching (on a single GPU), and resultant models can deploy to run real-time in browser-based applications on edge devices including tablets and smartphones. The compressed models provide comparable cutting-edge accuracy, while having a 30 times smaller model size and can run 40.7 ms per frame in a popular browser on a popular smartphone and OS.