Enhancing old photos with Deep Learning
- 0 Collaborators
Combining in the right way some Deep Neural Network we can enhance old photos non only restoring them but also improving the resolution and recreating more details. ...learn more
Project status: Concept
oneAPI, Artificial Intelligence
Groups
DeepLearning,
Artificial Intelligence Europe,
Intel AI DevCamps
Intel Technologies
DevCloud,
oneAPI,
DPC++,
OpenVINO,
Intel Python
Overview / Usage
Looking at old photos I started to think about how to recover them from scratches, holes, and how to improve the quality adding details.
Reading many research papers I found many interesting methods but no the right one to archive my goal.
For this reason, I decided to combine more approaches in order to create a hybrid solution able to reconstruct and improve old photos.
The solution it's based on a combination of a fully convolutional neural network (U-Net) and a Generative Adversarial Networks (GANs).
Methodology / Approach
Working with models that require a lot of memory we can face on out of GPU memory limitation. For example, the NVIDIA 2080 it's limited to 8GB and the expensive NVIDIA RTX8000 it's limited to 48GB. We know that an Intel i9 10th generation can utilize up to 128GB of memory and the Intel® UHD Graphics 630 can utilize up to 64GB of memory. An Intel Xeon W-3275M can utilize up to 2TB of memory!
So the main idea it's to utilize a technology that allows us to don't be limited by GPU memory for training and inference. This is one of the main reasons why I chose oneAPI as the core technology for this project. oneAPI defines several mechanisms for sharing memory across the platform (CPU+devices), depending on the capabilities of the devices (GPUs, FPGAs, other accelerators).