2023 Author: Bryan Walter | [email protected]. Last modified: 2023-05-21 22:24
Google developers have introduced NIMA - a deep convolutional neural network that determines the highest quality photos. The algorithm of its work, described in detail in the company's blog and preprint on arXiv.org, is based on two parameters: the technical component of the photograph and its general aesthetics, assessed by a person.
Modern technology allows you to take excellent quality pictures using your smartphone camera and related image processing applications. However, the true quality of a photograph is determined by its aesthetics, which consists in the correct composition and well-chosen lighting - parameters that are difficult to automatically detect. Previously, another Google neural network, Creatism, learned how to create aesthetically pleasing photos from images taken with Google Street View cameras.
A new neural network, created by Hossein Talebi and Peyman Milanfar of Google Research, will be able to identify the best photos from the ones that the user took himself: NIMA (Neural Image Assessment) is trained to choose from a series of images the one that the average user is a person would consider the highest quality and aesthetic. To train the neural network, the developers used the AVA (Aesthetic Visual Analysis) database, which contains about 200 thousand photographs, each of which was rated by professional photographers on a scale from 1 to 10. Researchers trained the neural network to guess the rating of an image based on its analysis: NIMA correctly guessed the score of professional photographers - people with an accuracy of 80 percent.
Sample images from the AVA database and ratings: NIMA and real (in brackets)
As a training sample, NIMA used a series of photographs of one object (for example, parrots) and altered brightness or clarity and ranked them by rating, thereby choosing the highest quality.
NIMA rating of different parrot images
According to the authors of the work, such a neural network can be used in modern smartphones: for example, choose the best from a series of photos taken by the user or determine which filter or color rendering settings are best suited.
Recently, developers from the University of Bern presented a neural network to improve image quality: an algorithm based on Bayesian deep learning can effectively correct even blurry images.
