Multiobjective Reinforcement Learning-Based Neural Architecture Search for Efficient Portrait Parsing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Efficient deployment of deep learning models on resource-constrained edge devices remains a challenge.
• Portrait parsing requires high accuracy and low latency for real-time applications.
• Existing neural architecture search (NAS) methods often struggle to balance performance with computational cost.

Purpose of the Study:

• To develop efficient portrait parsing models deployable on edge computing or terminal devices.
• To balance accuracy, model size (parameters), computational cost (FLOPs), and inference latency.
• To introduce a multiobjective reinforcement learning (RL)-based NAS scheme for optimizing these trade-offs.

Main Methods:

• Designed a multiobjective reinforcement learning (RL)-based neural architecture search (NAS) scheme.
• Employed a two-stage training strategy with precomputing and memory-resident feature maps.
• Incorporated knowledge distillation to accelerate convergence and improve the RL controller's signal.
• Searched for architectures on CelebAMask-HQ dataset.

Main Results:

• Generated Pareto-optimal architectures balancing accuracy, parameters, FLOPs, and latency.
• Achieved state-of-the-art performance: 96.5% MIOU on EG1800 (portrait segmentation) and 91.6% F1-score on HELEN (face labeling).
• Developed models with approximately 0.3M parameters, outperforming human-designed networks in accuracy while consuming fewer resources and offering higher real-time performance.

Conclusions:

• The proposed RL-based NAS scheme effectively finds efficient portrait parsing models for edge deployment.
• The developed models achieve superior accuracy and efficiency compared to existing methods.
• The approach demonstrates successful transferability to other portrait and face segmentation tasks.