Simplified Algorithm Of Manga Line Draft Based On Generative Adversarial Networks

The creation of comics and manga can be divided into three steps: sketch,line draft,and coloring.The process that a cartoonist draws a line draft from a sketch is called "line draft simplification".The line draft clearly outlines the various elements in the comics,such as characters,animals,etc.,which is the key to understanding the semantics of the comics.The quality of the line draft is crucial to the final effect of the comic.However,the comic drawing cycle is long,and almost all work can only be done manually,often requiring multiple comic professionals to work together,which takes a lot of time and energy.There are already some algorithms for studying line draft simplification,using algorithms to replace cartoonists to automatically complete the simplification process from sketch to line draft.Traditional methods usually use multiple edge detection operators or recognize lines and strokes based on rules and thresholds.There are more and more applications of the generative adversarial network.As a generative model,it has achieved good results in the field of image translation.Therefore,this paper mainly studies the manga line draft simplification algorithm based on generative adversarial networks.This article first introduces the basic concepts related to the simplified task of line drafts,then collects a large number of manga sketches and line drafts from multiple photo websites,and uses "line draft reverse processing" to increase the amount of alignment data,and finally builds the training The set contains two types of data: alignment data and individual line drafts.There are 19228 sketch line draft samples.This paper also builds a test set consisting of 129 aligned sketch-line draft pairs.Then this paper proposes a new evaluation standard,combining quantitative indicators and human observation to comprehensively evaluate the quality of the generated line draft.Finally,a simplified model of linear drafts based on multi-class data was constructed on the training set.The two baseline models pix2 pix and Cycle GAN were first trained with single-class data.The experimental results proved the importance and superiority of the alignment data;based on the baseline models,two or three types of data are used to train the simplified model of line drafts after the improvement of VGG network to reconstruct content loss and discriminator to use Patch GAN.By comparing the quantitative indicators of different models through experiments and analyzing the effect of generating linear manuscripts on the test set,we found that the simplified model based on alignment data and separate line drafts has the best simplification effect and the highest quality of the generated line draft,moreover this model can also generate high-quality line draft for the colored comic.