The new site of Let’s Enhance on the basis of neural networks raises the resolution and recreates the details of the uploaded image. Especially well the program copes with city landscapes and photos of animals. The review has published the portal PetaPixel.
Developers often use neural networks and machine learning to work with photos. Special algorithms allow you to remove noise, ” think out ” some details of the image and even deform the objects already present on it. In addition, neural networks are also used to improve image quality – but very often such programs are not available to ordinary users who do not have programming skills.
Programmers Alex Savsunenko and Vladislav Pranskevicius created an easy-to-use website. To improve the quality of the picture, just drag the image to the middle of the page. Further, the algorithm removes all artifacts, scales the image, increasing 4 times, and also dorisyvaet new details in the picture. The waiting time is only a few minutes.
The results of processing, in general, look good, the best program copes with photos of animals and buildings. So far, neural networks have definitely not been given work with portraits – people’s faces look painted. In addition, the algorithm has difficulties in processing small parts.
Programmers said that their system is inspired by the neural networks of SRGAN and EDSR . They used a large amount of teaching material, 10 percent of which were photographs of individuals. In the future, researchers hope to create separate neural networks for each type of image. First, the program will determine what is shown in the picture, and then apply the necessary tools.Neural networks also know how to transform images significantly. So, the French developer Grigory Antipov created a neural network, which can realistically change the age of the person pictured in the photo. Similar, but implemented much more simply feature recently provided by the application FaceApp.