Research On Chalk Drawing Style Simulation Based On Deep Convolutional Neural Network

Using of computers to simulate different styles of artwork is one of main directions in the field of non-photorealistic rendering.Researchers have successfully simulated some artistic styles such as oil painting,landscape painting and sketching in past two decades.The achievements have been widely applied in engineering drawing,medical medicine,animation production and other industries.As a type of painting with a long history,chalk painting known by its rough lines and brilliant colors.However,research on art of chalk drawing in the field of non-photorealistic rendering are not deep enough.In the field of non-photorealistic drawing,the mathematical modeling is often used to simulate works of art.This method relies on specific algorithms to process the images step by step in order to obtain predictable results.However,this simulation algorithm is too specific and can't be generalized to a large number of natural images usually.This thesis intends to combine learning and convolutional neural network to study the chalk style simulation algorithm.Deep convolutional neural network has made great research progress in computer vision since 2006.Different from the traditional algorithms based on mathematical modeling,convolution neural network model has a very strong generalization feature,just to make up for the weakness of traditional algorithms.Thesis proposes a chalk drawing style simulation algorithm mainly in four aspects of constructing,training,using and improving the chalky style network model.First of all,the construction of network model is divided into two parts: the chalkgenerating network and the chalk-discriminating network.The generating network is responsible for generating picture with chalk-style,and the discriminating network is responsible for determining whether the chalk-drawing style of generated pictures is up to standard.The chalk-generating network and the chalk-discriminating network are convolutional neural network.Secondly,training of the network model relies on loss function constructed on the chalk discriminating network.The loss function is used to calculate difference between generated picture and chalk drawing,generated picture and output image.The error information is fed back to the chalk drawing network,and the generating network update network parameters accordingly.When the network model converges,it is considered that model has been trained.Thirdly,the usage of network model is very simple.Input a picture into the chalkgenerating network,after a network operation,picture has the artistic style of chalk drawing.Finally,the network model proposed in this thesis improves on the formation of chalk-generating network,the construction of loss function,and the preprocessing of training images in order to obtain better experimental results.The experimental results show that the chalky style network model in this thesis can well simulate the characteristics of real chalk art,especially the rough sense of lines and color intertwined visual effects.Compared with traditional algorithms,the generation speed and algorithm generalization ability have been greatly improved.Broaden the research scope of non-photorealistic rendering technology.