Research And Application Of Word Replacement Based On Area And Style Change

With the development of social media,more and more people are engaged in picture creation.Text replacement is widely used in eliminating sensitive words,generating subtitles,producting expression images and so on.For this task,there are few researches on word level text replacement in wild scenes.The mainstream research schemes have some shortcomings,such as poor effect on distorted text replacement and rough texture of text details.In order to solve the above shortcomings,we propose ACNet(area change network)and DINet(decoupling integration network).The main contributions are shown below.Firstly,we propose a text replacment network ACNet based on area change.The network uses the detection model to detect the position of keypoints,and uses the thin plate spline interpolation network to distort the text area,so as to reduce the difficulty of style change.At the same time,the foreground and background separation model is proposed to separate the text and the background area,so that the model pays more attention to the text area and improves the text replacement effect.Secondly,we propose a text replacment network DINet based on text style feature decomposition and fusion.The network extracts local and global text features through local-global attention module.We use Residual Full Connection layer for feature decomposition,and Residual AdaIN model to fuse the decomposed features.Meanwhile we use area smoothing model to improve the effect of edge region replacement.On the dataset in this paper,DINet achieves better results than SRNet.Thirdly,we design and implement a text replacement platform,which is convenient for users to replace text.The usability of the platform is verified through experiments and tests.