Research On Visible Watermark Removal And Robust Extraction Algorithm Based On Deep Learning

Visible image watermarking is often used in image copyright protection of social platforms and digital galleries.With the development of Internet and self-media technology,a large number of digital images are stolen and embezzled,especially watermarks containing copyright marks on commercial images are often illegally removed.In order to improve the robustness and anti-removal ability of image visible watermark and explore the complete copyright extraction on the image attacked by removal,the security analysis of visible watermark removal is a key topic worthy of study.The security analysis of visible watermark is mainly divided into two parts:watermark removal and robust extraction.Watermark removal aims to remove the watermark from the image without damaging the content information of the image,and generate the corresponding watermark-free image,so as to verify the effectiveness and robustness of the watermark and provide reference for safe and reliable watermark design.The purpose of the robust extraction task is to mark the original watermark area by extracting the traces left by the removal operation on the stolen image,which can effectively obtain evidence when the works are infringed and safeguard their own interests.However,the current security analysis still has the following two problems:1)Most watermark removal algorithms have not considered the information fusion and content consistency of the removed area and surrounding areas,which leads to the appearance of watermark shape artifacts in the removal results and affects the overall removal effect;2)Because of the transparency of the watermark,the removal operation has more information to recover the original area under the watermark than other pixel tampering operations,and the traces left are relatively less obvious,which makes it difficult to extract.Focusing on the above two problems,this thesis proposes a new visible watermark removal network and robust extraction network,and the specific work is as follows:1)A visible watermark removal algorithm based on sparse mask is proposed.First of all,in the first stage,CBAM(Convolution Block Attention Module)attention mechanism and GMSD(Gradient Magnitude Similarity Deviation)loss are added on the basis of the existing multitasking codec to improve the mask segmentation accuracy and restoration results.In addition,a restoration network based on sparse mask is added as the second stage of the network,and some pixels in the restoration result of the first stage are regenerated from outside to inside to reduce the watermark contour artifact in the image.The two stages together constitute the watermark removal network,which ultimately makes the removed area and surrounding areas have higher content continuity and style consistency.Compared with the most advanced algorithms,the four self-made watermark datasets in this thesis are all improved,among which the watermark data sets with pure white and opacity between 0.3-0.7 are the best,which is improved by 0.51 compared with the latest algorithm PSNR,and other related indexes are also improved to some extent.The subjective test proves that the watermark removal result obtained in this thesis has higher naturalness.2)A watermark robust extraction algorithm based on NAS(Neural Architecture Search)and attention mechanism is proposed,which can effectively extract the copyright of watermarked images after watermark removal attacks.Referring to the recent image restoration detection network,the detection process is divided into three parts:enhancement block,extraction block and decision block.The enhancement block consists of several filter layers and preprocessing layers including Gaussian blur difference module,which are used to enhance and remove traces.The extraction block is generated by the latest neural search algorithm few-shot NAS search,which improves the generalization performance and extraction ability compared with manually designed convolution block.Finally,the decision block consists of cooperative attention mechanism and global attention mechanism,which are used to measure the importance of information from two angles of channel and location,so as to achieve better clustering.Finally,this thesis introduces a boundary loss to generate a mask with more complete contour.In this thesis,the generalization performance of the algorithm is studied on five self-made watermark removal data sets,and the data sets that can make the network achieve the best generalization are selected.At the same time,the performance of the algorithm is verified,in which AUC can reach 97.47%and F1 can reach 87.27.Compared with similar algorithms,the performance of this algorithm is better.