- Neural networks can automatically enhance your pictures without ever being trained on what a noise-free photo looks like.
- It can be used in low-light photography, magnetic resonance imaging and physically-based image synthesis.
Wouldn’t it be nice if you could take your low-light, pixelated or grainy photos and remove the artifacts and noise without using photoshop. A new machine learning model can do the same by simply observing samples of corrupted images.
Signal reconstruction from corrupted measurements is a crucial part of statistical data analysis. Nowadays, we are seeing a lot of interest in avoiding the conventional statistical modeling of signal corruptions, due to recent advances in machine learning techniques.
Researchers at MIT, NVIDIA and Aalto University have applied a basic statistical reasoning to signal reconstruction using neural networks. It learns to restore signals without ever looking at clean ones.
It’s different from other state-of-the-art methods or recently develop image-enhancing AI. While other machine learning techniques in this field focus on training a neural network to restore photos by showing both noisy and clean pictures, this method only requires a pair of input images with the grain or noise.
This artificial intelligence system can automatically enhance your pictures without ever being trained on what a noise-free photo looks like.
Conventional machine learning techniques involve training a regression model like a convolutional neural network with large datasets containing pairs of corrupted inputs (noisy images) and clean target (fixed image), and reducing the empirical risk.
On the other hand, in this method, the clean targets can be disposed completely, as long as the network is able to observe each source image twice. It can be trained to fix images with significant (50 percent) outlier content. Sometimes, it outperforms the model using clean exemplars. Moreover, it’s a less expensive task than obtaining the clean target.
To train the network on 50,000 images, the researchers used NVIDIA Tesla P100 GPUs with TensorFlow framework powered by CUDA deep neural network library.
There are numerous real-world scenarios where acquiring clean training data is a tough task: low-light imaging like astronomical photography, magnetic resonance imaging and physically-based image synthesis.
Obviously, the network can’t learn to pick up features that aren’t present in the input images, but the same is true for training with clean targets.
MRI reconstruction example | Courtesy of researchers
In this study, the researchers started with standard noise distribution (including Additive Gaussian noise), and continued to the tougher, analytically intractable Monte Carlo noise in image synthesis. They also observed that image reconstruction from sub-Nyquist spectral sampling in MRI (magnetic resonance imaging) can be learned from noisy images only.