一种基于 CGAN 的可见水印去除方案-《计算机技术与发展》

文章信息/Info

Title:: A Scheme of Visible Watermark Removal Method Based on Conditional Generative Adversarial Nets

Author(s):: WANG Jia-liang; LIU Xiao-qiang; LI Bai-yan; FENG Zhen-ni; School of Computer Science and Technology,Donghua University,Shanghai 201620,China

Keywords:: image conversion; watermark removal; feature points matching; conditional generative adversarial nets ( CGAN ) ; supervised learning

摘要:: 自然场景下采集的卡证、文档中存在的可见水印,是影响人们阅读效率、机器识别准确度的障碍。为此,提出了一种结合基于特征点匹配的水印检测和基于条件生成对抗网络 CGAN 的水印去除方案。水印检测部分,通过 SIFT 特征点检测、FLANN 特征点匹配和 PROSAC 误匹配消隐,估计出几何变换的最佳透视模型实现目标水印定位。水印去除部分采用了 pix2pix( image-to-image translation with conditional generative adversarial networks)的模型架构,它借鉴了 CGAN 的核心思想,混合了 L1 距离损失和 CGAN 损失函数,减少了输出图像的模糊度且保留了更多的正确特征。最终能满足自然业务场景下快速且精准的水印去除需求,具有较好的水印检测鲁棒性和去水印效果。此外,还给出了详细的 CGAN 模型所需的成对训练集扩充方式,构建了大量有效的训练集,提升了去水印模型训练的拟合优度。

Abstract:: Visible watermarks collected from cards and documents from natural scenes has been an obstacle to reading efficiency and recognition accuracy. To solve this problem,a new watermarking scheme is proposed,which combines watermark detection based on featurepoint matching and watermark removal based on conditional generation adversarial networks. In the watermark detection part, SIFT,FLANN and PROSAC are used to estimate the best perspective model of geometric transformation to realize the target watermarklocation. In the watermark removal part, pix2pix model architecture is adopted,which uses the core idea of CGAN for reference andmixes L1 distance loss and CGAN loss function to reduce the ambiguity of the output image and retain more correct features. Thisscheme can satisfy the requirement of watermark removal in real business scenarios with excellent robustness. In addition,the expansionof paired training set for CGAN model is introduced in detail. The large number of effective datasets constructed have greatly improvedthe goodness of fit of the CGAN model.