Face anti-spoofing based on multi-model features
As face anti-spoofing is becoming a more and more popular technology, it is important to protect face recognition systems from the attack. There are many available face anti-spoofing benchmark datasets used for face anti-spoofing this recently. In this thesis, I firstly introduce some basic knowl...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis-Master by Coursework |
| Language: | English |
| Published: |
Nanyang Technological University
2020
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/143107 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | As face anti-spoofing is becoming a more and more popular technology, it is important
to protect face recognition systems from the attack. There are many available face anti-spoofing
benchmark datasets used for face anti-spoofing this recently. In this thesis, I
firstly introduce some basic knowledge and developing history of face recognition
technology. At the moment, face recognition technology is wildly used in various
fields, such as railway security system, education and smart city construction.
However, face recognition systems currently are easy to be attacked by various of
methods, including photo attack, video clips and two dimension or three-dimension
mask, causing the recognition result more unreliable. In order to solve this problem,
face anti-spoofing becomes a hot topic. In this thesis, we describe various methods of
anti-spoofing methods, including Color Texture, Patch and Depth-Based CNNs, DMD
+ LBP, Pulse + texture, Deep Pulse and Depth, Micro-texture + SSD and De-Spoofing.
After that, we analyze and compare the numbers and characteristics of different face
anti-spoofing data sets. We do improvement to the multi-modal fusion method to better
combine these four chosen features: RGB, Depth, IR and HSV. They are modaldependent
features and are re-weighted in order to choose informative channel features
and at the same time suppress the useless ones. Finally, we conduct experiments on
CASIA-SURF which is current the most complete multi-modal dataset in the world.
The results show that the TPR of squeeze and excitation fusion method are 7.6%, 48.2%
and 39.0%, which is better than halfway fusion method with the FPR=0.01, 0.001 and
0.0001, respectively. |
|---|