Research On Landscape Painting Image Generation Based On Few Shot Of Unsupervised Learning

With the rapid development of computer vision,the emergence of generative adversarial networks improves the authenticity of synthetic images and generates richer data.At the same time,generative adversarial network is also widely used in landscape image generation tasks,such as natural scenery,Chinese painting image generation.In order to stimulate the author’s creative inspiration and improve the image quality,it is necessary to fuse different styles of images based on the author’s paintings to generate new high-quality images.However,the number of landscape images in real life is limited,and the design of generative adversarial model has great challenges due to its high requirements on texture,style and natural attributes in landscape painting creation.Traditional GAN image generation methods rely on a large amount of data,and when faced with Few-shot input,the generated images are not good in quality and diversity.Under unsupervised conditions,aiming at the characteristics of Few-shot art painting images,poor quality of generated images,and inflexible style,this paper proposes a new method of small sample art painting image generation,which integrates different styles of images to generate high-quality images.The main research contents of this paper are as follows:(1)In this paper,a new unsupervised Few-shot image generation method FUGAN is proposed to solve the problem of poor image quality caused by too few data under unsupervised conditions.Different styles of images are fused to generate new high-quality multi-style images.Content on the generator set image and image at the same time as the input of the learning style,and this strategy can be in the process of learning to create a direct feedback path,helps to choose the style influence each other,less in the sample data to make full use of the sample data,generate style flexible high-quality image.In addition,using multi-scale fusion learning strategies in the generator,the input image to study under the different resolutions and fusion,to strengthen the study of landscape painting different scales,from local to the global context,the details can be effective learning,effectively solve the problem of data dependency in the process of training,improve the learning ability of small sample data,generate high quality and style more images.(2)Aiming at the problem of local structure disharmony and ignoring global information in the process of art painting image generation,FUGAN-SA method is proposed on the basis of FUGAN: In the FUGAN-SA,the self-attention mechanism module is introduced to measure the weight ratio by calculating the pixel position relationship between the content image pixels,strengthen the correlation of image features,so as to give more weight,establish the dependency relationship between features with strong correlation,and improve the overall coordination of art painting images.In the setting of content encoder and image decoder,the residual block is used to replace the ordinary convolution layer to retain the inherent features of the image,reduce the information loss of the original image features in the middle convolution process,and restore the image details better,which can effectively improve the performance.A multi-task antagonistic discriminator is designed on the discriminator setting to increase self-reconstruction and reduce the change of irrelevant image domain information,further guide the direction of image generation and improve the effect of image generation.