Javascript must be enabled for the correct page display

Improving face segmentation using feature extractors fine-tuned on gender recognition

Lindström, Arvid (2018) Improving face segmentation using feature extractors fine-tuned on gender recognition. Bachelor's Thesis, Artificial Intelligence.


Download (872kB) | Preview
[img] Text
Restricted to Registered users only

Download (94kB)


This study investigates the benefit of fine-tuning encoders used in segmentation networks prior to training on semantic segmentation data. The domain of the study is face segmentation through pixel-wise labelling using three models: VGG16, VGG19, and ResNet-50 as encoders. A cross- comparison study is performed where encoders trained previously on Imagenet are compared with encoders trained on Imagenet followed by fine-tuning on a gender recognition task. The dataset used for gender recognition is the CelebA dataset. The datasets LFW and HELEN are used for face segmentation. It is demonstrated that segmentation networks built on VGG16 and VGG19 obtain an average IoU increase of 3.9% and 11.0% respectively when encoders are tuned on gender recognition prior to being used for face segmentation.

Item Type: Thesis (Bachelor's Thesis)
Supervisor name: Wiering, M.A.
Degree programme: Artificial Intelligence
Thesis type: Bachelor's Thesis
Language: English
Date Deposited: 25 Jul 2018
Last Modified: 27 Jul 2018 12:48

Actions (login required)

View Item View Item