Image Local Style Transfer Algorithm Based On Convolutional Neural Network

In modern social life,people have higher and higher requirements for images.For the problem that it is difficult to directly realize the style transfer of part of the image in practical application,an image local style transfer algorithm based on convolutional neural network is proposed.At present,image style transfer algorithms are mainly based on image optimization and model-based optimization.The first kind of algorithm takes all pixels in the image as parameters,and changes the value of pixels in the image through iterative optimization to make it conform to the characteristic distribution of style image.The second kind of algorithm trains the model to learn the feature information in the style image,Thus,the content image is directly mapped to the style transfer image.Based on the first kind of image style algorithm,this paper combines it with the image semantic segmentation algorithm to improve the original algorithm and realize the local style transfer of the image.The main work includes:In the part of image semantic segmentation algorithm,aiming at the problem that some low-level and middle-level details are lost when the existing semantic segmentation algorithms extract the high-level semantic information of feature map,a FW-deep lab v3 + image semantic segmentation algorithm based on atrous spatial pyramid pooling module and weighted fusion of feature images is proposed.The improved algorithm is based on the "residual idea",establish the jump connection between feature extraction network and ASPP module,weight and fuse the feature map,increase the ability of ASPP module to extract image semantic information,reduce the loss of detail information of low and middle-level parts,alleviate the problem of gradient disappearance,improve the segmentation performance and operation efficiency of the model.The experimental results show that the reasonable fusion of feature maps makes the edge details of semantic segmentation map better preserved,the segmentation accuracy of the improved algorithm is improved,and its average intersection and merging ratio is 6.8% higher than that of the original model,and the frequency weighted intersection and merging ratio is 5.33% higher than that of the original model.In the part of image style transfer algorithm,aiming at the problem that the style transfer of the specified area of the image can not be carried out directly in practical application,an image local style transfer algorithm based on mask constraint is proposed.With the help of FW-Deep Lab v3 + algorithm,the content image is segmented and its target area is extracted,using the image style migration algorithm optimized pixel by pixel based on Gram matrix,the image local style transfer content and style loss function are redefined,on the one hand,the mask matrix generated by semantic segmentation is used to restrict the range of parameters to be learned,on the other hand,the mask matrix is used to restrict some areas calculated by style loss function to remove redundant areas.Experimental results show that the improved algorithm has better local style conversion ability,improves the performance of the algorithm and speeds up the convergence speed of the algorithm.