Noise2noise, a Technology to Clear Pictures off Visual Distortions

What word would you choose to explain the haziness or grainy droplets in an otherwise perfect picture that you recently shot through your smartphone or camera? A team of researchers terms it as “noise” and claim to clear the visual distortions and pave way to enjoy a perfect picture. The team includes professors from the affiliation of MIT in the U.S., Aalto University in Finland, and NVIDIA.

According to a posting on the NVIDIA Developer News Center, the research is based on deep learning approach that has enabled fixation of photos merely taking a few samples of corrupted photos. It has been detected that to produce a noise-free photography, a long exposure is required. It has been learnt that with ideal circumstances, signals can be reconstructed after observing a few examples of unclear signals.

It must be noted that vision algorithms in computers are already utilizing tools to improve pictures that are taken from elite smartphones such as iPhone X and Pixel 2, but the research expects to take it to the next level. Moving over the usage of subjecting a pair of images to neural networks, in which one was blurry and other of high quality, the novel model called noise2noise can understand how to clear images without the needing the high resolution exmples.

The researchers have made use of NVIDIA Tesla P100 GPU with TensorFlow, which was accelerated by cuDNN, in order to formulate a deep learning network. The new technology promises to enhance magnetic resonance image (MRI) scans too!