Neural network turns scribbles into art masterpieces

Program Neural Doodle, made on the basis of a convolutional neural network is a script doodle.py, which generates an image, taking three or four images as input. Including input is fed a simple sketch (what the authors call "scribbles") and a sample of the style, with its outline. For example, in the case of the above example of the style pattern is such a pattern Renoir.

Neural network extracts the characteristic style features - and puts them on a sketch

. Here is another example.



A sample of the style of Claude Monet.



To run the program needs Python 3.4+, installed numpy and scipy library and python3-dev. pre-trained neural network is required for the script (VGG19, 80 MB). For installation instructions, see the local environment. Here.

For rendering on the GPU requires a good Nvidia card with CUDA technology, and 2-4 GB of memory (for larger images - 8-12 GB). Rendering can be run on the CPU, in this case, you need about the same amount of RAM.

The design of the neural network is described in the scientific work of the author of "Semantic Style Transfer and Turning Two-Bit Doodles into Fine Artworks", which he had prepared for the conference nucl.ai Conference 2016.

Neural network uses image synthesis algorithm, which is proposed by researchers Chuan Lee (Chuan Li) and Michael Vendome (Michael Wand) scientific work "Combining Markov Random Fields and Convolutional Neural Networks for Image Synthesis".

Source: geektimes.ru/post/272430/