APSIPA Transactions on Signal and Information Processing > Vol 12 > Issue 1

Boundary-Aware Face Alignment with Enhanced HourglassNet and Transformer

Yingxin Li, Shandong Provincial Key Laboratory of Network Based Intelligent Computing, University of Jinan and School of Information Science and Engineering, University of Jinan, China, Dongmei Niu, Shandong Provincial Key Laboratory of Network Based Intelligent Computing, University of Jinan and School of Information Science and Engineering, University of Jinan, China, ise_niudm@ujn.edu.cn , Jingliang Peng, Shandong Provincial Key Laboratory of Network Based Intelligent Computing, University of Jinan and School of Information Science and Engineering, University of Jinan, China, jingliap@gmail.com
 
Suggested Citation
Yingxin Li, Dongmei Niu and Jingliang Peng (2023), "Boundary-Aware Face Alignment with Enhanced HourglassNet and Transformer", APSIPA Transactions on Signal and Information Processing: Vol. 12: No. 1, e28. http://dx.doi.org/10.1561/116.00000115

Publication Date: 24 May 2023
© 2023 Y. Li, D. Niu and J. Peng
 
Subjects
 
Keywords
Face alignmentfacial landmark detectionboundary heatmapHourglassNet
 

Share

Open Access

This is published under the terms of CC BY-NC.

Downloaded: 751 times

In this article:
Introduction 
Related Works 
Proposed Approach 
Experiments 
Conclusion 
References 

Abstract

In this work, we propose a neural network for boundary-aware face alignment. The proposed network is composed of two stages with the first one estimating boundary heatmaps and the second one predicting landmark positions. We build the first stage by enhancing a baseline HourglassNet. Major enhancements include {the} addition of a CoordConv layer and addition of shallow and deep feature fusion (SDFusion) blocks. For the second stage, we design a subnet that firstly fuses information of the original image, a latent feature from the first stage and the boundary heatmap generated by the first stage, and secondly uses a Transformer to map the fused feature to the landmark coordinates. As shown by experiments, the proposed algorithm achieves state-of-the-art performance on the benchmark datasets.

DOI:10.1561/116.00000115