Engineers from NVIDIA, Finnish Aalto University and the Massachusetts Institute of Technology have collaborated to create a revolutionary self-learning technique for correcting noise-tainted photos.
Scientists wrote about their development back in March this year. Then they explained the mathematical principles behind the self-learning of a neural network. In July, the idea was presented at the ICML international computer learning conference in Stockholm.
The most common problem for photographers is that pictures taken in low light are often “noisy,” meaning they look grainy and washed out. Mobile photos taken with a budget device suffer especially strongly from this. In this case, AI can be of great help to the photographer, capable of restoring the image to an acceptable state.
The proposed idea is to teach a neural network to correct images by analyzing pairs of grainy and distinct objects. This requires showing the AI as many examples as possible. Subsequently, he will be able to recognize and restore snapshots on his own.
The team trained their system based on 50 thousand images taken from the ImageNet database. Positive results were achieved with NVIDIA Tesla P100 GPUs and Google powered cuDNN deep learning platform TensorFlow.
According to the members of the project, the development points the way to fundamentally new solutions for effective noise reduction.
The Topic of Article: NVIDIA AI can correct “noisy” photos.