Skip to content

Latest commit

 

History

History
68 lines (64 loc) · 4.72 KB

README.md

File metadata and controls

68 lines (64 loc) · 4.72 KB
Raw

awesome_face_antispoofing

  • This is a single shot face anti-spoofing project.
  • The deep learning framework is Pytorch. Python3.5 is used.

Installation

  • sudo -s
  • sh install_requirements.sh
  • exit

Face landmarks

  • face_alignment is used for landmarks extraction. Page face_alignment. Thanks to them.

Landmarks extraction scripts

  • cd detlandmark&&python3 detlandmark_imgs.py NUAA_raw_dir

Data

  • I have upload data and detected landmarks into GOOGLE DRIVE-raw.tar.gz

  • I have upload data and detected landmarks into Baidu DRIVE-raw.tar.gz

  • You can change corresponding directory and filename in config.py

  • For example train_filelists=[ ['raw/ClientRaw','raw/client_train_raw.txt',GENUINE], ['raw/ImposterRaw','imposter_train_raw.txt',ATTACK] ] test_filelists=[ ['raw/ClientRaw','raw/client_test_raw.txt',GENUINE], ['raw/ImposterRaw','raw/imposter_test_raw.txt',ATTACK] ]

Method

  • Our method is straightforward. Small patched containing a face is cropped with corresponding landmarks. A binary classification network is used to distinguish the attack patches. alt text

Training

  • First, edit file config.py, choose the target network and proper batch_size.
  • Then, in terminal command: make clean&&make&&python3 main.py train

Inference

  • In terminal command: python3 inference.py inference --images='detlandmark/inference_images/*/*.jpg'
  • The inference report is result/inference.txt, you can check it in commad: cat result/inference.txt

Visualize Dataset

  • We have fixed the bug of choice wrong face in multiple detected faces with standard of coordinates.
  • To visualize cropped faces in dataset. Please run command: python3 vis_cropface.py visualize
  • All faces will be shown in data/showcropface_train.jpg and data/showcropface_val.jpg
  • The training data are shown here. Training Validation

Experiments

  • Experiments results on NUAA[1] Image input size is as same as the imagenet.
  • State-of-the-art networks are used, e.g. VGG[2], ResNet[3], DenseNet[4], Inception[5], Xception[6], DetNet[7]
Network Acc AUC EER TPR(1.0%) TPR(.5%)
VGG-11 0.9416 0.99600562 0.031592 0.948099 0.931262
VGG-13 0.9452 0.99261419 0.034890 0.908696 0.869814
VGG-16 0.9591 0.99449404 0.027599 0.952283 0.926575
VGG-19 0.9013 0.99623176 0.023086 0.958378 0.941503
Res-18 0.9813 0.99872778 0.008158 0.992470 0.989585
Res-34 0.9656 0.99978646 0.003992 0.998091 0.996181
Res-50 0.8677 0.99951550 0.008923 0.991668 0.986544
denseNet121 0.9803 0.99872628 0.014754 0.981534 0.970144
denseNet161 0.9757 0.99610439 0.016664 0.977222 0.967020
denseNet169 0.9334 0.99532942 0.029744 0.949662 0.932130
denseNet201 0.9263 0.99833348 0.012195 0.985718 0.975525
InceptionV3 0.8078 0.99172718 0.036278 0.927270 0.907655
Xception 0.9843 0.99973281 0.005728 0.996431 0.993101
DetNet 0.9072 0.99998322 0.000892 0.999705 0.999703

Reference

  • [1]Tan X, Li Y, Liu J, et al. Face liveness detection from a single image with sparse low rank bilinear discriminative model[C]// European Conference on Computer Vision. Springer-Verlag, 2010:504-517.
  • [2]Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv:1409.1556, 2014.
  • [3]He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 770-778.
  • [4]Huang G, Liu Z, Van Der Maaten L, et al. Densely connected convolutional networks[C]//CVPR. 2017, 1(2): 3.
  • [5]Szegedy C, Vanhoucke V, Ioffe S, et al. Rethinking the inception architecture for computer vision[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 2818-2826.
  • [6]Chollet F. Xception: Deep learning with depthwise separable convolutions[J]. arXiv preprint, 2017: 1610.02357.
  • [7]Li Z, Peng C, Yu G, et al. DetNet: A Backbone network for Object Detection[J]. arXiv preprint arXiv:1804.06215, 2018.