📜- Da Vinci Algorithm

Image

An image, the base of a digital artwork, is defined on a finite set Ω of pixel locations. We denote by w and h the width and height of Ω, respectively. Due to the human visual perception principles, three-primary colors, namely Red Green and Blue (RGB), are sufficient for the description of a certain visible color. Thus, an image I is a mapping that I(p) = (R,G,B), where p denotes a pixel, R,G and B are integer values between [0,255]. From such definition it can be observed that an image, or say a digital artwork, can be construct from scratch by computing.

Da Vinci Artwork Generator Overview

Theoretically, a deliberate artwork I involves two components, i.e. content and style. Here content indicates the semantic information and the composition of I. The content is the backbone and soul of an artwork. The style defines the details of an artwork, e.g. color schemes or textures. Style leads to the formal aesthetics of an artwork.

Our Da Vinci Algorithm is an intelligent artwork generator which produce digital artworks based on state-of-the-art deep learning technology. Users manipulate the content of the output artworks by providing a set ΞC of content example images, and express style preference by another set ΞS of style example images.

Content Creation

We generate novel content by using Generative Adversarial Networks (GAN). A generator network is trained on the user-specified image set ΞC to create sample similar to the images in ΞC visually, but without repetition. An evaluating network is also trained to judge the output quality of the former network. Such two networks evolve together in adversarial. The resulting model, denoted by Δ(ΞC), generates high-quality novel images with content hints of ΞC.

Artistic Style Rendering

We employ Neural Style Transfer (NST) to give artistic style to an artwork. A properly trained multi-layer convolutional neural network plays the role of feature extractor. Fingerprint information of images in style image set ΞS is achieved by the neural network. We run an iterative optimisation process with dual target (C, S).

Da Vinci Algorithm Summary

In summary, our Da Vinci Algorithm can be formalised by I = Λ[Δ(ΞC), ΞS]. Both the two components mentioned above, i.e. Δ and Λ, take large computation. Benefit from the parallel computing nature of neural networks, this algorithm can be distributed. Generally speaking, the algorithm generates user desired digital artworks by using cutting-edge artificial neural networks deployed on large-scale distributed computing. A generated artwork, joints intention of users, is unique in content and elegant on artistic style.