Research On Detection And Removal Of Image Visible Watermark Based On Deep Learning

With the rapid development of Internet technology,various kinds of information have been rapidly and widely spread by means of the Internet.Image has become an important information carrier of Internet content because of its unique intuitiveness.A large number of images spread on the network make visible watermark technology widely used in image copyright protection.The existing conventional visible watermark detection and removal methods have many limitations in practical applications.In recent years,deep learning has made great progress in the field of image processing.Therefore,the development of visible watermark detection and removal methods based on deep learning to solve the problems existing in traditional methods has become more important in practice.This paper proposes a framework for detecting and removing image visible watermarks based on deep learning.The framework consists of two parts: the detection of visible watermarks and the removal of visible watermarks.(1)The detection part of the visible watermark: This paper regards the detection of visible watermark as target detection,and proposes a multi-branch adaptive decision-making visible watermark detection algorithm based on Faster R-CNN target detection algorithm.Specifically,first,when the visible watermark template is known,multiple visible watermark datasets are obtained by changing the strength of the visible watermark added to the image and the scale ratio of the visible watermark to the added image.Next these datasets are input into the Faster R-CNN network for training to obtain different branch models.Then these branch models are combined to construct a multi-branch adaptive decision-making visible watermark detection network,in which each branch is responsible for detecting a certain range of visible watermark with added strength and size.Finally,the adaptive decision maker in the detection network selects the final detection box and the visible watermark category by non-maximal suppression of the detection box and classifier score obtained by each branch.Due to the lack of visible watermark detection datasets for deep learning,this paper builds new visible watermark datasets and conducts a comprehensive experimental evaluation of the algorithm.The experimental results show that the visible watermark detection algorithm proposed in this paper can effectively detect visible watermarks with different positions,different added intensities and different sizes in the image,and the detection accuracy is improved compared with the Faster R-CNN algorithm.In addition,the paper also studies the influence of various factors on the visible watermark detection algorithm of multi-branch adaptive decision making.(2)The removal part of the visible watermark: This paper proposes a visible watermark removal algorithm based on conditional adversarial network.The visible watermark removal algorithm adopts the idea of image transformation to realize the removal of the visible watermark.The entire visible watermark removal network based on conditional adversarial consists of two generators and one discriminator.The first generator performs a preliminary removal of the visible watermark portion detected in the image.The conditional adversarial network composed of the second generator and discriminator is used to further remove the visible watermark and generate high-frequency information.In this paper,the visible watermark removal dataset is constructed and the visible watermark removal algorithm is evaluated.The experimental results show that the visible watermark removal algorithm proposed in this paper can effectively remove the visible watermark in the image and preserve the original image content under the visible watermark coverage.The purpose of the algorithm proposed in this paper is to hope that the design of the visible watermark should be able to effectively resist the attack of visible watermark detection and removal based on deep learning while highlighting the copyright statement.The visible watermark can better achieve the purpose of image copyright protection.