Source: ATYUN AI platform
Generating confrontation Network (GAN) has been used to discover new drugs, create compelling burgers and butterfly photos, and create a synthesis scan of brain cancer. The new paper Maastricht University (Maastricht University) published show that it does not generate a difference in terms of identity.
Research paper "LoGAN: Generating Logos with a Generative Adversarial Neural Network Conditioned on color", the scientists describe an AI system, called LoGAN, you can make 12 different colors of the logo.
The researchers said, "as any designer logo is a lengthy, complicated and expensive process. However, recent advances in generating algorithm provides a model can provide possible solutions, LoGAN results for the first time shows how AI for help designers creative process, and open the future direction of the promising, for example, include a more descriptive label, these labels will provide more detailed and easy to use system. "
The researchers explained that the arts generate GAN problem is that they do not always produce the results aesthetically pleasing. Their solution is to use color to define the most prominent logo: black, blue, brown, blue, gray, green, orange, pink, purple, red, white and yellow.
The team on the LDD-icons datasets training system, the data set consists of 486,777 small icon composition, size is 32 by 32 pixels. Each dominant color extraction algorithm used, and color conversion from RGB values to words. At the same time, the system of the third neural network (except generator and a discriminator) sample images are classified.
So how LoGAN effect? Although the resulting logo blurry (due to the low resolution of the source image), but some very convincing logo. When the input color keyword, it tries to obtain an irregular shape, round and square buttons, even Google Chromelogo appearance.
Interestingly, white and gray colors is one of the 12 categories of the most common of the three color combinations. Class orange, brown is the preferred neural network, and yellow in the class, it is usually in blue.
The researchers believe that, like AI system LoGAN can handle some of the more monotonous work logo design, designers were able to brainstorm. In future work, they want to expand the system to understand the semantics of the word beyond the color of keywords, such as the shape and focus.
They wrote that an improved system could be trained on two different sets of data: a logo containing a clear geometry, the other comprising a non-regular shape logo. It may use the embedded model, the most commonly used words describe logo to "improve the interpretability."
"Although the generated identity has a very low resolution, but they can serve as a rough draft of the final logo, or as a means of inspiration for designers," the researchers wrote, "if given a keyword model, can be successfully recommended create a logo, in our case, these keywords by the logo of the most prominent colors. such keywords can be considered to be descriptive, because it provides an easy-to-distinguish human logo property. "
Paper: arxiv.org/pdf/1810.10395.pdf
This switched ATYUN artificial intelligence media platforms, the original link: AI design of small expert: choose a color for the AI to help you generate logo
more recommendations
Brief series forecasting problems
Depth learning to change the use of location-aware computing
