Research On Cartoon And Caricature Generation Based On Style Transfer

As an important artistic style,cartoon combines the characteristics of the characters and the unique painting effects,and is gradually loved by people.With the development of image processing technology,the research on image is no longer confined to image recognition,target detection and tracking.The research on image generation and image-to-image translation has been carried out,and some progress has been made.However,in some existing methods,the method of processing an image is relatively complicated,and the obtained image has obvious synthetic traces with a simple style.Therefore,this paper aims to automatically generate cartoon faces using a deep learning-based approach to learn facial features and preserve the original content featuresThe main work of this paper are as follows:1.In this thesis,a method of generating cartoon faces based on feature fusion is studied.For the case of paired data,the L1 Loss function can be used directly between the generated image and the ground truth,making the generated image as close to real ones as possible.On the basis of generative adversarial network,this paper proposes to add a network to extract more representative facial features,in order to accelerate network convergence and improve the quality of generated images.At the same time,the face key-points are used to constraint the important parts of generated images,making the image details more realistic.2.In this thesis,a method of generating cartoon face based on content invariance is studied.Considering the content invariance of image in style transfer,a content feature extraction network and a style feature extraction network are constructed respectively to encode the image.The content feature integrated with different style feature can be decoded to images with different styles.In order to further improve the quality of generated images,this paper proposes a multi-scale discriminant network to discriminate image blocks of different sizes and combine their scores to gain the final discriminant result.The network can not only improve the discriminating ability.It can also make the generated image more realistic.3.In this thesis,a method of generating caricature based on key-point shifting is studied.The caricature generation involves both texture changes and shape deformation.Besides style transfer,the module of face exaggeration deformation is proposed to warp the images by automatically estimating face key-points.Finally,the spline interpolation is used in neural network for point-based warping.Because it is differentiable,this module can be trained as part of an end-to-end system with style transfer.The network can warp the photo into a caricature,while preserving identity.In order to verify the effectiveness of the algorithms,this paper constructs a special cartoon and caricature database.The method and all the comparison experiments are based on the same database for training and testing.The experimental results show that the quality of generated image by those methods has improved.