Classification according to artists and styles of painting, to humans is difficult; and in a variety of artists and styles of clear links between difficulty is greater. Therefore, if you let the machine do the work is definitely impossible, but is it so?
In academic research, there are few areas of computer science and machine learning there is no influence, one of which is the history of art. Even the most advanced algorithms, analysis of authors, content and style of painting is almost impossible.
However, with the development of machine learning techniques in recent years (such as depth Convolutional neural networks), this seems to be changing. As recently as a few years, computer scientists have developed the ability to match human, sometimes even beyond the human pattern recognition algorithms.
What genre is this a painting? Ask's algorithm
Babak Saleh and Ahmed Elgammal from Rutgers University in New Jersey, they make this technology a reality. The two people using the latest machine learning techniques, training algorithm for precise identification properties of painting, including works of artists and styles, which had never been done before.
More importantly, this technique can be used in the overall style of painting, identifying links between artists, many art historians usually takes years to understand that.
Saleh and Elgammal first created an image database, which includes over 15 centuries, more than 80,000 pieces of more than 1000 artists paintings. These covered 27 different styles of painting, includes more than 1500 painting examples within each style. Researchers also classifies the genre of the work including, city, landscape and so on.
, This paper studies staff will set the image subset, then take advantage of the subset to train different types of machine learning algorithms to identify the specified features paintings, including some common, lower-level features, such as the overall color of paintings; later training machine learning higher level features, such as depicted in the paintings of objects (like a horse, crosses). Eventually, the machine learning can be implemented with 400 different dimension vectors describe the paintings. Moschino case
Next, the researchers to identify a picture of the machine it has never seen the painting and the results have been impressive, the machines can accurately identify the artist, accuracy rate of up to 60%, know that artificial recognition accuracy is 45%. But most of all, machine learning techniques to provide a solution, you can better understand the characteristics of art of these human struggles.
Art history might be able to let the computer school
On the classification of paintings but, machine learning, unable to do the best for the moment.
To cite an example, Saleh and Elgammal machine learning algorithms is difficult to distinguish between kamier·bishaluo and Claude Monet's paintings. But artists need to do a little research can only clear distinction, late 19th century and early 20th century two people in France is very active, and Acad e Mie Suisse in Paris art college, using these backgrounds, experts know Camille Pissarro and Claude Monet are a pair of good friends and each other will also share their experiences of the arts. So, if their art works are somewhat similar, is not surprising.
https://www.youtube.com/watch?v=Y0KoC5Xf0Co
There is one example, machine learning or well differentiated between Monet and United States Impressionist painter caierde·hasamu, which by the France impressionist painters and Monet's impression. Similar to the Association, it needs to be judged.
In addition to the above-mentioned deficiencies, machines, it can look for links between the artistic style. For example, it will often confuse the abstract expression of attention and action painting because the artist on canvas through drip or dump paint paints to complete the work. Saleh and Elgammal said that if a human observer should be able to understand this confusion. They said, "the action can be seen as abstract expressionism in painting a seed type. "
Machine algorithms in similar paintings can also "find the difference." Like expressionist School of painting, and Fauvism, which would normally be considered to be a form of expressionist School of painting. In addition, machine algorithm can also be linked humanism and Renaissance art style, which clearly reflects the fact that moral values was integrated into in early Renaissance paintings.
Other links including, early Renaissance painting and Renaissance, Impressionist and Postimpressionist paintings, late Cubism and synthetic Cubist painting.
These links if you want to allow art historians to analyse it, may take decades or even hundreds of years, but by machine learning methods, just a few months. This is very important for the study of art history, the significance of new machine learning algorithms have an application is that you can pick out paintings with similar characteristics (see chart). It will be a very powerful tool, as historians may never know what links exist between different artists, and how different artists on works between interaction, and machine learning can explore more.
Machine learning algorithms provide a new form of artistic exploration, might bring unexpected results in the art world!
VIA technologyreview
1010 votes
Millet Note
Millet Note using the front panel 2.5D glass design, so the edges of the screen we can see a "film killers"-level arc. Using special 3D glasses behind the design, the two sides are arcs, the Visual effect is somewhat similar to the Samsung Galaxy Note Edge surface of the screen.
View details of the voting >>
No comments:
Post a Comment