Image Style Transformation Method Research Based On VGG And Local Labeling

Deep learning is a hot direction in the field of artificial intelligence.It has shown strong learning and processing abilities in image recognition,semantic segmentation,natural language processing and other fields,and has made great achievements,even surpassing human performance in some fields.Therefore,in recent years,many image processing researchers have used deep learning technology in image generation tasks and achieved good results in image style transformation.Image stylization is to transform the style of one image into another while keeping the content of the original image unchanged.Deep learning in the application of image style transformation,the use of neural network to complete the transformation of style can enable the machine to automatically generate creative images like those drawn by artists.However,there are still many problems to be solved in network training,generation time,stylization effect and other aspects of the current image style transformation method based on deep learning.In this paper,a comparative study is conducted between the image reconstruction method based on iterative optimization and the image reconstruction method based on the transformation network.In addition,local annotations are added into the iterative optimization method for region segmentation,and the corresponding regions are transformed respectively,so as to generate more realistic style images.This paper first analyzes the VGG19 model,studies the parameters of the VGG19 model,passes images through image-net,and visualizes each output feature layer of VGG19.Therefore,it is more intuitive to determine the specific representation of content features and style features.Then,input images aremarked locally through semantic segmentation,color maps are generated,and guidance channels are formed.Then,statistical feature extraction and statistical texture modeling are carried out respectively in the corresponding regions through the Gram matrix with guidance to complete the style transformation of local image reconstruction.Compared with the Gatys method,the Gatys method in this paper has the fastest speed and lower loss in terms of the gradient descent of the loss function.Thus,the conversion speed is faster and the image generated by the conversion is more accurate.From the perspective of image quality,the converted image generated by the method in this paper adopts local labeling,has no texture dislocation and is globally smooth without burrs.The images generated by other methods show obvious burrs and texture heterotopia.The method described in this paper has certain advantages.