AACNN

Tensorflow Inplementation for AACNN : Attribute Augmented Convolutional Neural Network for Face Hallucination (NTIRE2018)

Paper

Attribute Augmented Convolutional Neural Network for Face Hallucination
Cheng-Han Lee ¹, Kaipeng Zhang ¹, Hu-Cheng Lee ¹, Chia-Wen Cheng ², and Winston H. Hsu ^1
1 National Taiwan University, ² The University of Texas at Austin
IEEE Conference on Computer Vision and Pattern Recognition Workshop, (NTIRE 2018)
[Supplementary Material]

Visual Results

Dependencies

• Python 3.5+
• TensorFlow 1.8

Train the Model

• The Installation completely the same as our dependencies. Make sure you have correctly installed before using our code.
• Download the aligned version of CelebA dataset for training and testing
• Preprocess the training face images to 112 X 96, including detection, alignment, etc. Here we strongly recommend MTCNN, which is an effective and efficient open-source tool for face detection and alignment.
• Put aligned images under "./data/CelebA"
• For L2 version : bash train.sh #GPU
• For L2 + GAN version : bash train_gan.sh #GPU

Inference the Model

• bash test.sh #GPU
• For PSNR evaluation : run test_psnr.m on Matlab
• For SSIM evaluation : run test_ssim.m on Matlab

TO DO

• Add auxilary classifier loss
• Replace BatchNorm in discriminator with SpectralNorm