Research On Target Style Transfer Algorithm Based On Mask R-CNN

With the development of deep learning,image style transfer has quickly become one of the hot topics in the field of artificial intelligence.As an effective combination of computer science and art,style transfer algorithm endows computer art with the ability of re-creation.With the continuous innovation and iteration of image style transfer technology in the academic circle,its application in the industry has been constantly practiced and expanded,showing great potential in commercial value and artistic creation.At present,the style transfer technology is mainly oriented to the whole image transfer,while the target style transfer of the local area of the image has some application requirements,but relatively few studies.As a sub-task of style transfer task,target style transfer has different characteristics and requirements from global style transfer.Target style transfer can be decomposed into three sub-tasks:target extraction,style transfer and image fusion.The target extraction task needs to extract the target object from the image,including its classification information,location information and spatial layout information.The style transfer task is consistent with the global style transfer.The image fusion task needs to embed the generated specific stylized image into the original image according to the original positioning information of the target object,and uses the image fusion algorithm to process the embedded area,so as to make the fusion smoother and more natural.The existing research is carried out independently in various fields,and the research on integration is relatively rare.In the implementation of specific target style transfer task,the balance between speed and effect needs to be well handled,so as to improve the efficiency of algorithm processing on the condition that the final processing effect conforms to the expected condition.According to the technical requirements of each sub-task in the target style transfer,this paper implements the target style transfer algorithm based on Mask R-CNN.The main work of this paper is as follows:(1)By reviewing the classical target detection algorithms,such as R-CNN,Fast R-CNN and Faster R-CNN,the core principles and algorithm flow of the selected Mask R-CNN algoritlim framework are explained in detail.Through a brief introduction to the Keras and TensorFlow learning framework,the code implementation of Mask R-CNN and its key steps are discussed.It verifies the effectiveness of the Mask r-cnn algorithm in quickly and effectively extracting the target object's category information,location information and spatial layout pixel information from the input image.(2)Aiming at the target object's style transfer task,this paper introduces the realization of using the style transfer algorithn to transform the target object area in the content image into an image with a specific style.By summarizing the basic principle and algorithm flow of several key steps of image style transfer algorithm,a more efficient and fast style transfer algorithm is introduced.This paper introduces the generation style transfer model based on MS COCO data set and the forward inference of clipped content image.Finally,the experimental results are presented and analyzed preliminarily.(3)Aiming at the task of image style fusion,this paper proposes a two-stage image fusion scheme by analyzing the different effects of different statistical factors on the fusion effect.After initial fusion,this paper uses the model with fine visual adjustment to improve the fusion effect.This paper introduces in detail the initial fusion with certain robustness by matching style statistics and then improves the quality of image style reconstruction by re-designing the standard of style vector transmission and modifying the calculation method of loss function.Finally,the experimental results are presented and analyzed preliminarily.Experimental results show that the proposed image fusion algorithm can successfully embed and fuse the target region.