Research On Partial Image Style Transfer Based On Deep Learning

With the continuous development of artificial intelligence,image style transfer has gradually become a popular research content of deep learning.The image style transfer links computer technology and artistic creation together,so that computers can also perform artistic creation.The development of style transfer technology has made the application of this technology in the industrial field more and more frequent.Image style transfer technology has broad prospects and development space in terms of commercial value and artistic creation.Starting from the non-photorealistic rendering technology,which is the origin of the style transfer technology,this technology is carried out on a whole picture.Although there are many needs for style transfer only for certain areas of the picture,there are still few studies on this aspect.This paper introduces the semantic segmentation algorithm into the application of image style transfer,and at the same time constructs a network structure more suitable for style transfer.An algorithm that can perform style transfer on some areas of the picture is realized,and it is better than traditional algorithms in terms of the effect of individual style transfer.To complete this task,two steps,semantic segmentation and style transfer,need to be completed.The task of semantic segmentation needs to extract the target object from the picture.So we first build Mask R-CNN,use deep residual network and feature pyramid network as the model’s backbone feature extraction network,and then dynamically generate suggestion boxes through the regional recommendation network,and finally complete the target detection,classification and mask image generation.And after training on our labeled simple object data set,the task of extracting the objects of style transfer separately is completed.The style transfer task is to transfer the image style of the target object The traditional style transfer algorithm partly uses the VGG19 network pre-trained on the Imagenet dataset,and there is no network specifically used to implement style transfer.After doing a lot of style transfer experiments on several convolutional neural network models commonly used in the field of deep learning,this paper constructs a network model suitable for image style transfer.Compared with other pre-trained convolutional neural network models,the migration results are clear and not too distorted,and the number of parameters is reduced by 95%,and the running time of the algorithm is reduced by more than 22%.Secondly,the style loss function part of style transfer is improved,which can transfer multiple different painting styles for a content image at the same time.Figure 44 Table 5 Reference 65...