This repo implements the idea from this paper: https://pureadmin.qub.ac.uk/ws/portalfiles/portal/174233508/Attention_Based_Two_Stream_Convolutional_Networks_for_Face_Spoofing_Detection.pdf. The core motive behind this paper was to create a model which would be insensitive to illumination in detecting spoofing attacks.

This repo will be explained in 5 sections:

1. Data
2. CNN Backbone
3. Retinex Function
4. Train
5. Inference

Training and testing images are stored in test and train folders respectively. The folder will be organised as such: videos

train
    └── real
    	    └── real1.jpg
    	    .
    	    .
    └── fake
    	    └── fake1.jpg
    	    .
    	    .
test
    └── real
    └── fake

As shown in the image below, a CNN is needed in the MSR and RGB stream. I have decided to proceed with MobileNetV3 taken from https://github.com/xiaochus/MobileNetV3 due to its lightweight architecture.

This backbone can be substituted with any of the backbones here: https://keras.io/api/applications/.

The Multi-Scale Retinex (MSR) functions have been taken from: https://github.com/dongb5/Retinex. The core idea behind the use of MSR is written in the paper:

MSR can separate an image to illumination component and reflectance component, and the illuminationremoved reflectance component is used for liveness detection

To carry out training, a custom DataGenerator is needed as well as the Attention layer used in the paper.

The custom DataGenerator is found in datagen.py and it generates batches of images in the format [BxHxWxC RGB image, BxHxWxC MSR image] as shown in the image below.

The Attention-based Fusion stated in the paper can be found in attention.py. The inner working of the layer can be seen in the image below.

To test the model using your webcam, run the following command: python test.py