HAL will be down for maintenance from Friday, June 10 at 4pm through Monday, June 13 at 9am. More information
Skip to Main content Skip to Navigation
Journal articles

Two-stage human hair segmentation in the wild using deep shape prior

Yongzhe Yan 1, 2 Stefan Duffner 3 Xavier Naturel Anthony Berthelier 1, 2 Christophe Garcia 3 Christophe Blanc 1, 2 Thierry Chateau 2
1 COMSEE - COMputers that SEE
ISPR - Image, système de perception, robotique
3 imagine - Extraction de Caractéristiques et Identification
LIRIS - Laboratoire d'InfoRmatique en Image et Systèmes d'information
Abstract : Human hair is a crucial biometric characteristic with rich color and texture information. In this paper, we propose a novel hair segmentation approach integrating a deep shape prior into a carefully designed two-stage Fully Convolutional Neural Network (FCNN) pipeline. First, we utilize a FCNN with an Atrous Spatial Pyramid Pooling (ASPP) module to train a human hair shape prior based on a specific distance transform. In the second stage, we combine the hair shape prior and the original image to form the input of a symmetric encoder-decoder FCNN with a border refinement module to get the final hair segmentation output. Both quantitative and qualitative results show that our method achieves state-of-the-art performance on the LFW-Part and Figaro1k datasets.
Document type :
Journal articles
Complete list of metadata

Cited literature [50 references]  Display  Hide  Download

Contributor : Yongzhe Yan Connect in order to contact the contributor
Submitted on : Monday, July 6, 2020 - 2:16:09 PM
Last modification on : Thursday, September 9, 2021 - 2:36:02 PM
Long-term archiving on: : Monday, September 28, 2020 - 2:18:20 PM


PRLETTERS-D-19-00788R2-P5-12 (...
Files produced by the author(s)



Yongzhe Yan, Stefan Duffner, Xavier Naturel, Anthony Berthelier, Christophe Garcia, et al.. Two-stage human hair segmentation in the wild using deep shape prior. Pattern Recognition Letters, Elsevier, 2020, 136, pp.293-300. ⟨10.1016/j.patrec.2020.06.014⟩. ⟨hal-02890593⟩



Record views


Files downloads