Your phone may be able to clean the clichés – but our AI is much better to retouch, say the boffins • The registry



[ad_1]

Video Do not worry if the lighting is a little off in your photos – an artificially intelligent software can fix it.

Computer scientists from Nvidia, Aalto University in Finland and the Mbadachusetts Institute of Technology USA, formed a network of neurons to restore images spoiled by noise patches. Computer vision algorithms are already automatically used to improve shooting on smartphones like Pixel 2 or iPhone X, but that goes further.

With this new technique, the training process is slightly different from Google and Apple.

Instead of feeding the neural networks of a pair of images, where one is of high quality and the other is fuzzy, the latter model – nicknamed noise2noise – can learn to clean the images without the need for "We apply a basic statistical reasoning to signal reconstruction by machine learning – learn to map corrupt observations to clean signals – with a simple and powerful conclusion: in some common circumstances, it is possible to learn to restore signals without ever observing ones that are clean, "according to the document's summary.

The theoretical basis of why this works is a bit difficult to understand. and. The more traditional techniques use low resolution and high resolution image pairs, learn to minimize the loss function by estimating the difference in pixel values ​​between the two images.

Pixels can take a wide range of values ​​to recreate an image. image sharper, and over time, the neural network learns to average these values. The same idea can be applied when training on spoiled picture pairs, though the difference between the pixel values ​​in the two pictures is relatively similar to those between a clean and fuzzy picture.

"This implies that we can, in principle, corrupt the training targets of a neural network with zero noise without changing what the network learns," the paper writes.

Training with Technology

The team trained its noise2noise model on 50,000 images from the ImageNet dataset. a random distribution of noise to each image. The system must estimate the magnitude of the noise on the photo and remove it.

It has been tested on three sets of data with images of buildings, people and medical resonance imaging. Here is a video with some results.

Youtube video

The model will not cure all imperfections, however. It can not restore objects just out of frame or reposition the photo to get the best angles. But it is useful when there are not enough examples of high resolution to train like good images of galaxies or planets. "There are several real-world situations where getting clean driving data is difficult: low-light photography," our proof-of-concept demonstrations pave the way for significant potential benefits in these applications by eliminating the need from a potentially painstaking collection of own data.Of course, there is no free lunch – we can not learn to enter features that are not present in the input data – but this also applies to training with clean targets. "

The research is presented at the International Learning in Sweden Conference this week. MD

Sponsored by:
Minds Mastering Machines – Call for Papers Now Open

[ad_2]
Source link