AI programme rubs flaws off photos but doesn't know what clean images look like

American visual computing company Nvidia, along with researchers from Aalto University and MIT, have come up with an artificial intelligence (AI) program that removes defects from images such as grain, text and watermarks, without studying how clean photos look like.

“Our AI program shows significant benefits that can be reaped by removing the need for potentially strenuous collection of clean data (images),” the team said.

The AI program is expected to find use in the medical field to eliminate extensive post-processing that is deployed to remove noise from MRI (magnetic resonance imaging) images. “There are several real-world situations where obtaining clean training data are difficult: Low-light photography (for example, astronomical imaging) and MRI,” the team said. 

Previously, a neural network could cut out noise from images only after being trained on example pairs of unclean and clean images. But the new AI program has been exclusively trained on corrupted photos only. This means this AI program can remove noise from a photo without being trained on clean or noise-less images. 

Using Nvidia Tesla P100 GPUs with the cuDNN-accelerated TensorFlow deep learning framework, the team trained their system on 50,000 images in the ImageNet validation set.

The AI program is the latest example of how the emerging technology is being used in imaging.

Last month, Massachusetts Institute of Technology (MIT) announced an AI system that can analyse medical images in less than a second as compared with conventional techniques that may take two hours or more.

Under the conventional technique, for example, a brain scan from several months ago is put onto a more recent scan in order to reveal changes in a tumour’s progress. This process, under which traditional systems combine each pixel of one scan with the corresponding pixel of the other scan, can often be time-consuming, taking two hours or more.

The new system, however, does the task more than 1,000 times faster using the “learning” technique. Under this technique, the system “learns” how to align scans in the best and the quickest way by combining thousands of pairs of images. 

In April, TechCircle reported that travel marketplace Ixigo is using AI to scan photos of hotels. According to Ixigo, which is run by Gurugram-based Le Travenues Technology Pvt. Ltd, the AI engine can scan photographs of a hotel to find out if it has amenities such as a children's park or not.