American technology firm Nvidia along with researchers from Aalto University and the Massachusetts Institute of Technology have published a paper outlining a new artificial intelligence (AI) system that can almost flawlessly retouch grainy or pixelated pictures. The software uses deep-learning to automatically remove noise and artifacts.
No clean inputs required
What makes this system so unique is that it can teach itself to fix corrupted photos simply by looking at them. Previous deep learning work in image retouching was centered on training neural networks to restore images by comparing noisy and clear images.
This new method, however, only requires two corrupted images to proceed with removing noise. “It is possible to learn to restore signals without ever observing clean ones, at performance sometimes exceeding training using clean exemplars,” reads the paper.
The software will come in handy in the many real life situations and fields where clean images are simply not available or attainable. It will also mean that, in the future, photographers will have to worry less about creating ideal picture-taking conditions such as optimum lighting.
“There are several real-world situations where obtaining clean training data is difficult: low-light photography (e.g., astronomical imaging), physically-based rendering, and magnetic resonance imaging,” reads the paper's discussion section. “Our proof-of-concept demonstrations point the way to significant potential benefits in these applications by removing the need for potentially strenuous collection of clean data.”
Perhaps, the system's best asset is that it can perform faster, sometimes rendering frames in just 7 minutes, and as well or better than professional photo restorers. “[The system] is on par with state-of-the-art methods that make use of clean examples — using precisely the same training methodology, and often without appreciable drawbacks in training time or performance," reads the paper.
Applications in medical imaging
The researchers dedicated a special section of their study to Magnetic Resonance Imaging (MRI). This medical field is one of those applications that can particularly benefit from a software that can forgo the need for clean images.
"Many recent MRI techniques attempt to increase apparent resolution by, for example, generative adversarial networks (GAN) (Quan et al., 2017). However, in terms of PSNR, our results quite closely match their reported results," concludes the study's MRI section.
The system does have limitations. The researchers point out that it can not yet detect elements unavailable in the input photos.
However, the same drawbacks apply to softwares that employ clean inputs. "Of course, there is no free lunch – we cannot learn to pick up features that are not there in the input data – but this applies equally to training with clean targets," reads the paper.
The team's neural network was trained on 50,000 images in the ImageNet validation set using the NVIDIA Tesla P100 GPUs with the cuDNN-accelerated TensorFlow deep learning framework. It was then further validated on three different datasets.
The work will be presented today at the International Conference on Machine Learning in Stockholm, Sweden this week.