Chinese Document Watermarking Scheme Based On Automatic Generation Of Fonts

With the rapid development of computer network technology and the increasingly rich multimedia technology,the copy and flow of information become more and more simple,which also increases the risk of data leakage.Document is one of the most efficient and direct means of information recording,so preventing document leakage has always been a hot issue.Digital watermarking is one of the important means to solve the problem of document leakage.Through the document watermarking technology,the watermark information can be embedded into the document in a subtle but robust way.Without affecting the use value of the document,it can effectively control and track the source of leakage of electronic and paper documents.At present,the typical document watermarking methods can be divided into document structure based,natural language processing based,grid pattern based,image based and character geometry based methods.Among them,the document watermarking method based on character geometry uses the geometric characteristics of the characters in the document to express the watermark information,and achieves a good compromise in the visual quality,robustness and capacity.However,in the traditional Chinese document watermarking methods based on character geometry,the font library needs to be generated manually in the generation stage of the watermarked document,which greatly affects the efficiency of the watermarking.In order to make full use of the geometric characteristics of the characters and automatically generate the font library,so as to reduce the complexity of the Chinese document watermarking schemes,this dissertation designs robust document watermarking methods for two scenarios.The research work of this dissertation is summarized as follows.1)Chinese document watermarking method in digital transmission channelAiming at the distortion of common digital transmission channels(e.g.,screenshots,scaling,Gaussian noise and JPEG compression),this dissertation designs an automatic font generation based robust document watermarking scheme.The scheme composed of an end-to-end encoder-decoder structure based automatic font generation network,a character selection based embedder and a neural network based extractor.The font generation network can automatically complete the generation of the deformed font library,which is further utilized for embedding the watermark.In order to resist the possible distortion in the transmission process,a differentiable noise layer is added between the encoder and the decoder to simulate the distortion,so that the generated characters can obtain the corresponding robustness.The results show that compared with the existing Chinese document watermarking methods based on character geometry,the proposed method realizes the automatic generation of watermarked font library,and is superior to the existing methods in PSNR,SSIM and subjective quality score,which verifies the advantages of the proposed method in visual quality.At the same time,the proposed method can also effectively guarantee the robustness against digital channel distortion.2)Chinese document watermarking method in cross-media transmission channelMost of the existing document watermarking methods are still aiming at digital distortion.In this dissertation,a robust Chinese document watermarking scheme is designed for common cross-media channel distortion(e.g.,print-scanning and print-shooting).Based on the above watermarking scheme for digital channels,the scheme is optimized according to the characteristics of cross-media distortion.Because the cross-media noise layer is not differentiable,the above scheme cannot be directly used for the cross-media distortion.Therefore,this dissertation proposes a new noise layer design method to simulate the distortion caused by the print-scanning and print-shooting process,so that the watermarking method can obtain the robustness of cross-media transmission.At the same time,this dissertation proposes a two-stage training method,and fine-tunes the decoder to improve the watermark extraction performance.In addition,in the pre-processing stage of watermark extraction,a document character cutting scheme is designed for the print-scanning and print-shooting distortion,to ensure the synchronization of watermark information in the document.The results show that the watermark extraction rate of the proposed method is higher than that of the existing methods in the font size of three,four,small four and five.At the same time,the proposed method also has good performance in the scene of mismatched printing and scanning resolutions,which verifies the advantages of the proposed method in the print-scanning scene.For the print-shooting scene,this method achieves more stable watermark extraction performance than other methods under different font sizes.