Research On Chinese Landscape Painting Generation Algorithms Based On Multi-granularity Label Learning

The simulation research of Chinese landscape painting is one of the most challenging subjects in the field of computer graphics and images.In the past,the simulation work based on traditional models mostly focused on the simulation of the landscape brush method and local objects.Recently,there is a document based on the deep model of Generative Adversarial Network that has realized the translation from semantic layout label map to Chinese landscape painting.The composition and overall artistic style of Chinese landscape paintings are simulated,but this type of method has the problems of color and semantic distortion,large amount of network structure parameters,and the way for user interaction is too single.To deal with these problems,this paper studies the translation process of semantic layout label map to simulated Chinese landscape painting here,and proposes a multi-granularity label learning algorithm of Chinese landscape painting simulation generation.The main work of this paper is as follows:(1)A multi-semantic label map for landscape painting is designed as an interactive method,and a layered segmentation algorithm for generating this multi-semantic label map from hand-drawn landscape painting is proposed.Using a layered idea,the objects in the landscape painting are classified into 6 categories according to the three semantic levels of content,technique,and color,and designing the label map which corresponds color blocks to the multi-semantic content.This kind of multi-semantic label map is designed according to the specific characteristics of landscape paintings,which facilitates the interaction between users and the network;subsequent label maps are required for network training.For this reason,based on image processing technologies such as SLIC and Gray-Level Co-occurrence Matrix,a hierarchical segmentation algorithm for generating multi-semantic label maps from original hand-drawn landscape paintings is proposed.(2)Propose a CGAN-based algorithm for generating local color controllable landscape paintings from multi-semantic label maps.The main work of the algorithm is to design a lightweight Multi-Scale Color Class Concerned Conditional Generation Adversarial Network(MS3C-CGAN).The SPatially-Adaptive(DE)normalization residual block,Bilinear upsampling structure are introduced to simplify and reconstruct the existing UC-Net generator.The improved generator not only reduces the parameter amount by 24.45%,but also increases the simulation and control of colors on the basis of the original generation effect of landscape painting,and the semantic content is more accurate.(3)A granular progressive double-layer conditional generation confrontation network is designed to realize the generation of graffiti label maps to simulated Chinese landscape paintings.Considering that the multi-semantic label map is a fine-grained tag map,which is suitable for users with art foundation and not suitable for general users.Therefore,an algorithm for generating simulated landscape paintings in a graffiti interactive way is designed.The algorithm designed the first-level label map translation network C2FL-CGAN to translate coarse-grained graffiti label maps into fine-grained label maps,so that it has richer texture and layout information;the second layer introduces the trained MS3C-CGAN,and uses the fine-grained label map obtained in the first layer to generate the final simulation landscape painting.The experimental results show the feasibility and effectiveness of the method.(4)An automatic and rapid method for generating fine-grained label maps from landscape paintings to graffiti label maps is proposed.Since the C2FL-CGAN network training requires paired "fine-grained label map-graffiti label map" data,but the existing fine-grained label maps of landscape painting lack their corresponding graffiti label maps.Therefore,graphics techniques such as convex hull algorithm and DouglasPeucker algorithm are introduced to design a method of generating graffiti label maps from fine-grained label maps of landscape painting.