The Research Of Visible Watermark Removal Network Based On Decomposition Model

With the development and popularization of the Internet,people's awareness of copyright protection has become stronger.Visible watermarks are widely used to protect copyrights.Research on watermark removal effectively can provide enlightenment for the invention of more robust watermarking image technology.In addition,visible watermark affects image quality and visual effects,and is not conducive to certain basic computer vision tasks,such as text detection and recognition,image segmentation and other tasks.In view of this,the research on visible watermark removal is urgently needed.The changing characteristics of the shape,size,position,and transparency of the visible watermark bring difficulties to watermark removal task.Some existing mainstream watermark removal methods use convolutional neural networks to realize image-to-image conversion and translation,and have achieved good removal result.However,the watermark datasets used by these methods mostly contain gray-scale watermarks,which do not meet the needs of practical applications.Moreover,these methods don't have strong interpretability and can't separate the watermark from the original image.In this paper,some researches on the task of watermark removal are carried out as follows.(1)The dataset is an important factor to help train an excellent network.Aiming at solving the problems of the existing watermark dataset LVW,this paper constructs a Colored Large-scale Watermark Dataset(CLWD),which mainly contains colored watermarks to fill the gaps of current colored watermark dataset.After doing some experiments,it proves that the CLWD is superiority and challenging compared to LVW.(2)In order to realize the separation of watermark and non-watermark image,enhance the interpretability of the network and improve the it's performance,this paper combines the traditional watermarked image decomposition model to design a two-stage generator Watermark-Decomposition Network(WDNet).The first stage predicts the rough decomposition of the entire watermark image,and the second stage takes the watermark area as the center and refine the removal effect at the pixel-level.It shows that the separated watermark can help amplify the dataset to further improve the network's performance,and ultimately achieve the purpose of lifelong learning.Quantity experiments on LVW and CLWD consistently show that the WDNet is superior to the state-of-the-art methods in terms of accuracy and efficiency.(3)This paper found that the watermark is difficult to be recognized and removed when it has similar color with the background image.According to this,this paper starts from the two aspects of watermark position selection and transparency design to provide directions for inventing more robust watermark image construction technology,making the watermark in the image more difficult to be remove.