The Research Of Image Style Transfer Based On Neural Network

Style transfer is a highly creative field in computer vision.As deep learning returns to the academia,a large number of neural network structures are used in this field.Although excellent algorithms are continuously proposed,there is still huge room for improvement to efficient,robust and generalized model.In the process of local style transfer,the difficulty is how to balance the semantic content information and style feature information of images to obtain the results efficiently.In the process of global style transfer,the difficulty is how to integrate and utilize effective information in different levels of feature space to deal with multi-domain style transfer.For these two issues,this paper proposes two improved style transfer algorithms based on neural network.Previous algorithms based on neural network mostly aim at global style transfer,while some local style transfer methods are complicated and time-consuming.This paper proposes a GAN-based model(Painter GAN)that introduces self-attention mechanism and U-net.They make full use of multi-level image feature information to establish the correlation in global and local image respectively.Through adversarial training,the content image distribution is promoted to be closer to the target style image distribution.Experiments show that the algorithm in this paper can generate images in real time with higher efficiency and comparable or even better image quality.It shows that this local style transfer algorithm can well balance the realistic style and accurate content in generation.In view of the fact that most existing algorithms lack the generalization ability when dealing with multi-domain style transfer,the generated image lack diversity.This paper proposes a global and fast multi-domain style transfer algorithm.The algorithm classifies the target style according to the texture features,and adjusts important feature editing methods in terms of the classification to modify the images in the feature space.The model reconstructs the image with the edited feature maps in order from rough to detailed,and cyclic inputs at different levels.At the same time,in order to save more details,low-level and high-level reconstructed images are connected and output.Experimental results show that the algorithm in this paper can quickly complete multidomain style transfer,and generate images with better details and more artistic style.They prove that it is an effective strategy to solve the problem of multi-domain style transfer from the texture characteristics of the target style,which helps to improve the performance of the model.