Neural Network Methods For Multi-style Chinese Art Paintings Generation

Traditional Chinese paintings continue to innovate on the basis of inheritance.Among them,meticulous painting and ink painting are always the mainstream of Chinese art painting.In traditional Chinese painting,the description of flowers is an important and classic form of expression.Therefore,the article mainly carried out a series of research work on the intelligent creation of flowers in traditional Chinese painting.The meticulous paintings realize the simulation of painting objects through lots of colors and precise strokes,based on the same precise strokes as line drawing.The ink paintings pay more attention to the change of ink color,and the harmony between intensity and dryness,which achieve freehand expression by limited color and natural smooth strokes.From the methods based on traditional machine learning to the methods based on recently deep learning,image generation is always a hot area that scholars are constantly exploring.Style transfer is an important form of image generation,but it has some limitations,the features of using style image are obvious,the effect of generated image is single.With the rapid development of deep learning,generative adversarial network has become a popular method of image generation,which generates many impressive high-quality images through adversarial training between generator and discriminator.In numerous studies of image generation,however,very little research on the traditional Chinese art painting generation.On one hand,it is difficult to collect the dataset,on the other hand,due to the basic challenge of freehand expression in Chinese art painting styles and the difficulty of abstract expression of ink painting,existing methods cannot accurately generate the style and content of the Chinese painting.Aiming at these problems,we have carried out the task of multi-style Chinese art painting generation and the ink painting generation,based on generative adversarial network method,this article designs different network models to achieve effective expression of the two tasks,and achieves better results than other methods.The main contributions of this article are as follows:1.We construct an unpaired flowers dataset containing three classic traditional painting styles: line drawing,meticulous and ink.We have preprocessed them and they can be accurately used for training and testing.2.We propose a generative adversarial network model Flower-GAN based on attention mechanism to generate multi-style Chinese painting of flowers.By learning the mapping relationships between line drawing,meticulous and ink,the model finally obtains meticulous painting and ink painting images with certain style characteristics.By introducing the multi-scale structural similarity loss into the model,the high-frequency information of the source image is retained,the style consistency and content consistency can be accurately generated by combining it with adversarial loss,cycle consistency loss.3.We propose a model method based on adversarial neural network to generate ink paintings with unique style characteristics.This method consists of three steps in total.First,we propose to use the Line GAN model to learn the mapping from ink domain to line drawing domain,so as to generate line drawing images that retain more ink strokes and texture features.Second,data augmentation.The generated line drawing images are mixed with the real line drawing images to form a new dataset,and the morphological smoothing operation is performed on the dataset.Third,we propose to use the Ink GAN model to generate artistic ink painting images with a unique style,the introduction of total variation loss function in the model improves the spatial smoothness of the generated images.