Research And Application On Image Style Transfer Method Based On Encoder-Decoder

With the popularity of smart devices,the development of image applications has become increasingly prosperous,especially the image creation applications.However,most of these programs still stay in the functions of filter,pattern decoration and local beautification,and do not bring more creative experience.With the development of deep learning,image style transfer technology stimulates users' creative enthusiasm.Users can independently combine the target image and any style image to get the characteristic composite image.In order to deal with different creation scenes,image style transfer technology is divided into artistic style transfer and photo style transfer.Artistic style transfer has been welcomed in areas such as comic coloring and artistic style creation,while photo style transfer focuses on realistic photos,and it is commonly used in the field of game,film and animation.However,both artistic style transfer and photo style transfer have the problem of low quality in stylized images.In view of the problems,a study on the promotion of stylized image quality in image style transfer is carried out.First of all,the related knowledge of style transfer is introduced in this thesis.In addition,the algorithm related to style transfer is studied,and the attention mechanism and meta learning technology are briefly studied,which provides a theoretical basis for improving the quality of synthetic image in image style transfer technology.Aiming at the problem of low quality of stylized images in artistic style transfer,the synthetic images generated by the existing algorithms still have unclear content structure,the style of stylized images and style images are inconsistent in color,texture and shape.Therefore a method based on dual attention and meta learning is proposed in the paper.Firstly,the Ada IN is used to fuse content features and style features.A sub-dual attention module is constructed to enhance the expression of important features in fusion features from position and channel.Then,a meta learning system based on the encoder-decoder is built to strength the generalization of the model,improving the consistency of style.The experimental results show that the method improves the definition of the content structure,and to enhance the expressiveness of style characteristics,improve the quality of the stylized images.In view of the semantic information loss,texture distortion and deformation in the stylized images,a photo style transfer based on WCT and texture loss is proposed.Firstly,a multi-layer style transfer network is constructed,and the output of each layer encoder is transformed into content and style features by using WCT algorithm,aiming at fusing multi-scale information of different layers as feature expression,so as to make the generated image full and rich.Secondly,the Un Pooling operation is embedded in the shallow layer of the decoder,and the Unsampling operateion is embedded in the deep layer of the decoder to selectively restore the nodes with important features.Finally,texture loss and global style loss are used to construct the total style loss.The texture loss is based on the context information,while the global style loss is expressed by Gram matrix,which makes the stylized image maintain the global consistency of style,and also helps to retain the local details of style texture.Experimental results verifie the effectiveness of the proposed method.