Recent advances in machine learning such as deep convolutional neural networks are enabling the development of machines that perform pattern-recognition tasks.
Rutgers University researchers Babak Saleh and Ahmed Elgammal used machine-learning techniques to train algorithms to recognize the artist and style of fine-art paintings. The researchers began with a database of images of more than 80,000 paintings by more than a 1,000 artists spanning 15 centuries and covering 27 styles. They took a subset of the images and used them to train machine-learning algorithms to identify certain features, such as overall color and describing the objects in the image. The end result is a vector-like description of each painting that contains 400 different dimensions.
The researchers tested the algorithms on a set of unfamiliar paintings and found they achieved 60-percent accuracy in identifying the artist in the paintings it saw and 45-percent accuracy in identifying the style. The machine-learning approach also can pick up similarities, such as linking expressionism and fauvism, within just a few months. A potential application of the new algorithms is identifying paintings with similar characteristics to detect potential influences between artists.
From Technology Review
View Full Article
Abstracts Copyright © 2015 Information Inc., Bethesda, Maryland, USA
No entries found