Multi-task adversarial autoencoder network for face alignment in the wild

doi:10.1016/j.neucom.2021.01.027

Multi-task adversarial autoencoder network for face alignment in the wild

2021-05-01

9

PRE

AI

DOI：10.1016/j.neucom.2021.01.027

Original

Original request for help

Save

Abstract

En 中文

Face alignment has been applied widely in the field of computer vision, which is still a very challenging task for the existence of large pose, partial occlusion, and illumination, etc. The method based on deep regression neural network has achieved the most advanced performance in the field of face alignment in recent years, and how to learn more representative facial appearance is the key to face alignment. Based on the idea of Multi-task Learning, we propose a Multi-task Adversarial Autoencoder (MTAAE) net -work, which can learn more representative facial appearance for heatmap regression and improve the performance of face alignment in the wild. MTAAE is composed of three tasks. The main task uses the heatmap regression method to locate the position of landmarks and introduces a discriminator on the landmark heatmaps to generate more realistic heatmaps. Facial attribute estimation tasks and face recon-struction task based on Adversarial Autoencoder respectively extract discriminative and generative rep-resentations to improve the effect of heatmap regression. At the same time, the dynamic weight network is designed to assign a weight coefficient dynamically and reasonably for each auxiliary task. Extensive experiments on 300 W, MTFL, and WFLW datasets demonstrate that our method is more robust in com-plex environments and outperforms state-of-the-art methods. (c) 2021 Elsevier B.V. All rights reserved. Face alignment has been applied widely in the field of computer vision, which is still a very challenging task for the existence of large pose, partial occlusion, and illumination, etc. The method based on deep regression neural network has achieved the most advanced performance in the field of face alignment in recent years, and how to learn more representative facial appearance is the key to face alignment. Based on the idea of Multi-task Learning, we propose a Multi-task Adversarial Autoencoder (MTAAE) network, which can learn more representative facial appearance for heatmap regression and improve the performance of face alignment in the wild. MTAAE is composed of three tasks. The main task uses the heatmap regression method to locate the position of landmarks and introduces a discriminator on the landmark heatmaps to generate more realistic heatmaps. Facial attribute estimation tasks and face reconstruction task based on Adversarial Autoencoder respectively extract discriminative and generative representations to improve the effect of heatmap regression. At the same time, the dynamic weight network is designed to assign a weight coefficient dynamically and reasonably for each auxiliary task. Extensive experiments on 300 W, MTFL, and WFLW datasets demonstrate that our method is more robust in complex environments and outperforms state-of-the-art methods.

Keywords:

Multi-task learning

Adversarial autoencoder

Heatmap regression

Face alignment

Multi-task adversarial autoencoder network for face alignment in the wild

Abstract

AI Summary

Journal

Organization

Cited Papers