Research Of Manga Sketch Colorization Based On Deep Neural Network

As the primary form of animation industry on the market,manga has the characteristics of fast update and dissemination of the latest plots.In various animation agencies and anime fans,it is often only possible to obtain the newe st works in black and white hand-drawn versions.Color matching is one of the most important ways to enhance expressiveness of animation,and sometimes even the factors that determine the success or failure of the work,but it cannot be displayed because t he work will be update to the market with fast speed and colorization is time-consuming and need great effort.How to automatically color the black and white hand-drawn images with fast speed and nice effects,it is the automatic colorization technology of manga sketch.It has research and application value that is worthy of discussion and solution by researchers.Based on the deep learning theory,this paper proposes two kinds of manga sketch automatic coloring method based on deep neural network,which has achieved good results.The main content is as follows:1.Firstly,a method for manga sketch random colorization based on conditional generative adversarial network is proposed.By downsampling,the network extracts advanced features that can be used to segment regions,etc.,as well as low-level features such as color.In the upsampling process,according to the advanced features learned,the color distribution of the dataset and the random noise coloring corresponding regions.Aiming at the problems of color overflow and uncontrollable coloring of random coloring methods,an improved interactive coloring method based on conditional generative adversarial network is proposed.Using interactive color input as a constraint,when the network colors each area according to advanced features,it needs to fill the area with the corresponding color according to the color information from the interactive input in the area,thereby achieving the purpose of color control and solving the color overflow problem.In order to make the coloring effect and speed better,a series of optimizations are also applied to the network,such as using rescale plus convolution instead of deconvolution to solve the checkerboard effect,and using inversed convolution block instead of single convolution to learn more feature information and improve inference speed,etc.2.In order to implement the above two coloring methods,10,000 pieces of colorful manga image data are captured from the manga material website Safebooru using crawler technology.The Gaussian filter is used to suppress the noise point,and then the adaptive threshold is used to processing color image to obtain a manga sketch.Some image processing technique is used to generate a dataset that simulates the interactive color brush input effect.In order to verify the practicability of the proposed algorithm,a manga sktech coloring application was finally implemented based on Web technology.The application mainly includes an image data acquisition module,an image preprocessing m odule,a random coloring module,and an interactive coloring module.