Research On Watermark Area Localization Based On Image Segmentation

Image watermarking technology has become an important technical method for copyright protection and traceability of leaks by hiding secret information in the image carrier invisibly.However,there is a problem of insufficient robustness in practical applications,especially vulnerable to geometric attacks,such as rotation,scaling,cropping,and taking pictures.Since the geometric attack will cause the dislocation of the watermark area,the accurate localization of the watermark area is the premise of the watermark information extraction.Geometry correction of the watermark area by means of the location point or the edge of the image is an effective solution at present,but there are the following problems:(1)Methods based on location points or block effects rely on visually visible auxiliary location information,which affects the transparency and security of watermarks;(2)Using the edge of the image or the edge of the screen to locate the watermark area cannot resist cropping attacks.Therefore,how to realize the location of the local watermark area that is completely invisible and without any reference information has become the main goal of the research.Aiming at this goal,the deep learning image segmentation technology is innovatively applied to the watermark area localization problem by using the texture feature difference between the watermark area and the watermark-free area,and a watermark area localization method without reference information based on deep learning image segmentation is proposed.Starting from the imperceptibility and location morphological characteristics of watermarks,a watermark texture enhancement module and a spatial pyramid attention module are designed in the segmentation network to enhance the watermark texture features and capture watermark information at different scales.The loss function is also studied.The Dice loss function and the cross entropy loss function are combined to integrate their advantages.A combined loss function based on label smoothing and regularization as well as an auxiliary loss function is proposed.The experimental results show that the average intersection ratio and average pixel accuracy of the watermark localization method reach 99.02% and 99.41% respectively,surpassing the current mainstream segmentation networks Deep Labv3 Plus,DANet and PSPNet,etc.,and can complete the localization task of image watermark area with high precision.The proposed method has good robustness to geometric attacks and still has a high localization accuracy when the images are attacked.On the basis of the proposed algorithm,a watermark image localization system with B/S architecture is designed.The system greatly improves the practicability of the algorithm model by combining algorithm theory with engineering practice.After demand analysis,the modules of user management,model training,watermark area positioning and model management are designed,and the system is implemented by Flask and Spring Boot frameworks.