Design And Implementation Of Animation Sketch Automatic Colorization Algorithm Based On Gan Network

In the process of animation painting creation,the changes of character costumes and scenes are often not great,but the posture,action of characters and the angle of scene often change,so animation colorization is the most repetitive and time-consuming work content.In recent years,the rapid development of deep learning has provided a new automatic colorization method for coloring the animation sketch.This thesis studies the problem of automatic colorization sketches in the animation field,combined with the analysis of the artists’ needs and the difficulties in the animation sketch automatic colorization,Three colorization models are designed and implemented,achieved good colorization effects.In the existing Automatic Image Colorization Module,there are some problems,such as small amount of data and poor quality of sketch,which lead to unsatisfactory colorization effect;the existing Reference-Style Auxiliary Colorization Module lack the corresponding style graph of animation colorization,which leads to the module may not be strong enough to represent the style information and integrate it into the colorization process;the existing UserGuided Auxiliary Colorization Module lack the user-guided color information and cannot not guide sketch colorization effective.In this thesis,the following solutions are designed to solve the above problems,and their effectiveness is verified by experimental results:(1)In the Automatic Image Colorization Model,this paper designs and implements the sketch preprocessing module to deal with the problems such as too shallow contour in the line draft.In view of the problem that there are few data samples in the animation field and it is easy to over-fitting,this paper designs a memory enhanced colorization module to train the unsupervised representation of animation pictures,and embeds the representation in the network structure,designed based on Gan,to improve the colorization effect.The experiment shows that the model can still complete a good colorization effect in the case of few samples.(2)This thesis designs and implements the Reference-Style Auxiliary Colorization Model.In view of the disadvantage that the style information cannot be effectively integrated in the colorization process,a feature transfer module is designed in the reference attention mechanism in the model structure,which can fully learn the style information of the reference style graph and integrate it into the colorization process.Aiming at the lack of corresponding style graph in the training process,a self reference style graph generation algorithm is designed to simulate the style input in the real colorization process.In the part of loss function,this thesis introduces weighted style loss and content loss in the field of reference style transfer,and distinguishes the optimization goal of style information from the optimization goal of colorization content information.The experimental results show that the model can effectively capture the style of reference graph and color the sketch with similar style.(3)Although the Automatic Image Colorization Model and Reference-Style Auxiliary Colorization Model can obtain good colorization effect,the colorization result may not be consistent with the painter’s expected.Therefore,this thesis further designs and implements the User-Guided Auxiliary Model,and completes the sketch colorization meets the painter’s expectation by introducing the user’s color prompt.This thesis designs a color prompt generation algorithm for simulation,which solves the difficulty of lacking real color prompt in model training process.By introducing SE-Res NET into original model structure,the identification degree of each channel in feature representation is improved and the final colorization result is effectively constrained.