Neural Architecture Search for Portrait Parsing

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Neural Architecture Search (NAS) has shown success in image classification and NLP but is underexplored for dense-per-pixel prediction tasks.
• Current state-of-the-art methods for portrait segmentation and face labeling are manually designed, often lacking an optimal balance between performance and efficiency.

Purpose of the Study:

• To adapt and apply NAS methods to dense-per-pixel prediction tasks, specifically for portrait parsing, segmentation, and face labeling.
• To develop an efficient and high-performing model for portrait-related computer vision tasks using automated machine learning.

Main Methods:

• A cell-based encoder-decoder architecture was employed, featuring a carefully designed connectivity structure and search space for NAS.
• The proposed NAS method was applied to optimize models for portrait segmentation, face labeling, and portrait parsing tasks.

Main Results:

• Achieved state-of-the-art performance on three benchmark datasets: 96.8% MIOU on EG1800 (portrait segmentation), 91.2% F1-score on HELEN (face labeling), and 95.1% F1-score on CelebAMask-HQ (portrait parsing).
• The developed model demonstrates high performance with a low parameter count (2.29M), outperforming previous works.
• Empirically validated that NAS can significantly enhance even fundamental encoder-decoder architectures for portrait tasks.

Conclusions:

• The proposed NAS method successfully extends automated machine learning to portrait parsing, segmentation, and face labeling.
• This work represents the first application of NAS to these specific portrait-related dense-per-pixel prediction tasks, demonstrating its potential.
• The findings suggest NAS is a powerful tool for advancing computer vision in specialized domains like portrait analysis.