Research On The Improvement And Realization Of Style Transfer Algorithm Based On Deep Learning

The third upsurge of artificial intelligence comes from the rise of deep learning.From computer vision to driverless technology,deep learning technology has been applied to hundreds of practical problems by researchers.Especially in the field of image processing,the collision between artificial intelligence and neural art has attracted the attention of experts and scholars in related technical fields and art fields.This technology will change the way people process images.Style transfer refers to the use of deep learning technology to learn the style,color,shade and other elements of one or more style images with the same content information of a content image,so as to generate a kind of image with a famous painting style.According to the analysis of the current research situation,there are still many shortcomings in the application of deep learning technology to achieve image stylization.For example,after image stylization,the content information of the image has some problems such as unclear outline,color confusion,poor processing of content details and so on.The style information of the image has some problems such as distortion of texture details,unclear shape of strokes and so on.In order to solve the above problems,this paper studies the image style transfer technology as follows:(1)The algorithm of image style migration is improved and implemented.First of all,adjust the weight of image stylization style,the number of iterations,pooling method,target optimization algorithm,and so on.Experiment to determine the network super parameters of image stylization based on vgg-19 network.Then,select the methods of convolution layer content feature fusion and style feature fusion,and experiment to observe the influence of different fusion methods on image stylization effect.Finally,improve the average feature of sum The algorithm of style transfer is studied.The results show that vgg-19 model has the best image stylization effect with style weight,1500 iterations,mean pooling and Adam optimization algorithm.The convolution layer of content extraction is block3,the weight coefficient of content convolution layer is 1,and the convolution layer of style extraction is conv1?1?conv2?1?conv3?1?conv4?1?conv5?1.When the weight coefficients are 0,0,0.4,0.3 and 0.3 respectively,the effect of image stylization is better.The results of style transfer with style feature weights of 0,0,0.4,0.3 and 0.3 and content feature weights of 1 were compared with the results of image(2)stylization with average weighted concatenation of feature summation of each convolution layer.The experimental data showed that the result of image stylization with average weighted concatenation of feature summation of each convolution layer was better than that of single feature graph concatenation of convolution layer The PSNR and SSIM of the former were 0.91 and 0.95 higher than those of the latter.(2)Local style transfer algorithm based on deep learning is proposed.RTST network can't transfer the local style of image.Based on RTST model,the mask structure of image segmentation is introduced to build a new MRTST model of image style transfer network.Using RTST model and MRTST model to carry out style transfer experiments on images respectively,the image generated by MRTST model is 1.05 higher than the subjective evaluation MOS score of the image generated by RTST model,which obviously shows that the MRTST model proposed in this paper can effectively achieve image local style transfer,and the subjective evaluation score is high.