Research On Image Style Transfer Method Based On Attention And Feature Distribution Matching

Given a content image and a style image,style transfer refers to transferring the style patterns(tone,texture,or line features)of the style image into the content image,while enabling the content image to maintain its structure.In recent years,with the vigorous development of deep learning methods in the field of image processing and underlying vision,in the direction of style transfer,the method based on deep learning for style transfer has gradually been deeply studied and widely used.Based on the attention mechanism and optimal transport theory,this paper focuses on the style transfer task of images,aiming at the balance problem between content preservation and style feature transfer difficult to achieve by existing methods and the problem that may produce significant image artifacts.The problem is analyzed from two levels: 1)The current methods do not fully exploit the correlation between multi-level features;2)Some current neural style transfer methods mainly use CNN(Convolutional Neural Network)to extract low-level(mean and variance,etc.)statistical information of image features,so they cannot fully extract style from CNN features.Aiming at these two problems,this paper proposes an arbitrary style transfer solution based on a crossattention style transfer network model and a feature distribution matching method based on Wasserstein distance respectively,and achieves good experimental results.The research work of this paper is mainly divided into the following levels:(1)Balanced Image Style Transfer Based on Multi-level Normalized Cross Attention AlignmentIn this work,we first adopt the VGG-19 network to extract the multi-level features of the input content image and style image,and then extract the content features matching to the style features in different levels through multi-level content-style cross attention.In addition,through adopting the multi-level dynamic normalization and attention alignment,hierarchical content-style cross attention can effectively transform the content image through the style characteristics of different levels while preserving its local structure and semantics as much as possible.Apart from this,in the training phase,we also introduce a contextual loss function to supervise model training,which further improves the performance of the model.(2)Image style transfer based on style feature distribution matchingIn this work,we redefine the style loss as the Wasserstein distance between feature distributions based on optimal transport theory so that the style information can be more fully captured.And the GAN(Generative Adversarial Networks)structure is introduced here,the generator is a style attention model,and the discriminator is a Wasserstein discriminant model.We first train the discriminator,which is used to approximate the Wasserstein distance between the original style distribution and the generated distribution.Then we define a new style loss based on the output of the discriminator,and after this,we utilize the proposed Wasserstein style loss and traditional content perceptual loss to supervise the training of the generator together.Experimental results show that the stylized images generated based on this method have superior results compared with some existing typical methods.