ssjsusie/ipu-net

Identity-Preserved U-Net for Low Resolution Face Recognition

IPU-Net

We propose an Identity-Preserved U-Net which is capable of super-resolving very low resolution faces to their high resolution counterparts while preserving identity-related information. We achieve this by training a U-Net with a combination of a reconstruction and an identity-preserving loss, on multi-scale low resolution conditions.

Important Links

Paper
Project page
Blog post

Model Architecture

Our proposed method includes two deep neural networks which are designed to produce super-resolution faces from the lowresolution counterparts that are optimized for both visual quality and face recognition. The first network is an autoencoder with skip-connections (U-Net) and the other network is a pre-trained face recognition model with an Inception-ResNet architecture. The IPU-Net generates the super-resolved (SR) faces from the upsampled low-resolution versions by learning the mapping between the pairs. During training, the FRNet produces discriminative, low-dimensional embeddings for both SR and high-resolution (HR) faces in order to compute the identity loss.

Requirements:

Python 3.7
TensorFlow 2.0+
Pillow
Scikit-learn
OpenCV

How to train

To Train the model, run this:

python train.py

Result