视频可见水印检测与去除关键技术研究

Research on key technologies of video visible watermark detection and removal

针对可见水印多样性的特点和视频可见水印检测与去除过程中需要大量人工干预的问题,提出了一种基于图像修复的视频可见水印去除方法。该方法主要分为水印检测和水印去除2部分,水印检测采用目标检测和图像分割结合的方法,使检测的水印图案更准确;水印去除采用基于深度图像先验的图像修复实现,并在损失函数中引入结构相似度,修复网络通过学习水印以外的区域获得图像先验修复图像。为了验证该方法的有效性,用公开网站上的真实视频进行了综合的实验评估,实验结果表明该方法与其他修复方法相比,修复后的视频帧结构相似度和峰值信噪比更高。

Considering the diversity of visible watermark,a video visible watermark removal method based on image inpainting is proposed to solve the problem of large amount of manual intervention in the process of video visible watermark detection and removal.It is mainly divided into two parts:watermark detection and watermark removal.Target detection and image segmentation are combined to detect watermark patterns more accurately.Watermark removal is realized by image inpainting based on deep image prior,introducing structural similarity into loss function,the inpainting network acquires inpainting image by learning regions other than watermark.The experimental results show that compared with other image inpainting methods,the video frame repaired by the proposed method has higher structure similarity and peak signal to noise ratio.