With the rapid development of image acquisition equipment and multimedia display technology, recapturing high-quality images from various display medium becomes rather convenient. Such recaptured images have serious negative effects on image credibility and intelligent recognition systems. In order to solve the classification uncertainty problem introduced by the variety of the left artifacts of recapturing process, this project takes the left artifacts of recapturing process as the research object. By modeling the extraction process of local binary patterns (LBP) in deep learning network, the ability of LBP for describing the aliasing effect can be improved. Secondly, by constructing multi-scale deep color feature mining mechanism, the discrimination ability of traditional color feature can be improved. Lastly, with the help of resolution adaptive deep network, the influence of image resolution for accuracy can be suppressed. By complementing the advantages of traditional descriptive features of the left artifacts of the recapturing process and the advanced deep learning technology, the overall performance of recaptured image detection technology can be improved from the basic theory and the practical application levels. As a result, this research can not only provide theoretical basis and technical support for the further research of the recaptured image detection technology but also promote the popularization and application of recaptured image detection technology in many fields such as public security and intelligent recognition.
随着图像获取设备和多媒体显示技术的迅速发展,对显示在媒介上的图像进行拍摄从而获得高质量的翻拍图像变得愈发容易,这些翻拍图像对图像的可信度和智能识别系统的安全性有着严重的负面影响。针对翻拍过程遗留效应多样性所带来的分类不确定性问题,本项目以翻拍过程遗留效应为研究对象,通过在深度网络中对局部二值模式特征的提取过程进行建模,提升其对混叠效应的描述能力;通过构建多尺度深度颜色特征挖掘机制,提升传统颜色特征的判别能力;通过建立分辨率自适应深度网络,降低图像分辨率对检测精度带来的影响。实现传统翻拍过程遗留效应描述特征与前沿深度学习技术的优势互补,从基础理论和实际应用两个层面提高翻拍图像检测技术的整体性能,为翻拍图像检测技术的进一步研究提供理论基础和技术支撑,促进该技术在公共安全和智能识别等多个领域的推广应用。
随着数字图像获取设备和多媒体显示技术的迅速发展,对显示在媒介上的图像进行拍摄从而获得高质量的翻拍图像变得愈发容易,这些翻拍图像对图像的可信度和智能识别系统的安全性有着严重的负面影响。针对翻拍过程遗留效应多样性所带来的分类不确定性问题,本项目以翻拍过程遗留效应为研究对象,以图像处理和机器学习等学科为研究基础,从传统手工设计特征入手展开研究,随后借助深度学习技术进行端到端的训练和分类,最终实现传统翻拍过程遗留效应描述特征与前沿深度学习技术的优势互补。我们首先在深度网络中实现了纹理描述特征提取,提升其对混叠效应的描述能力;其次建立了颜色通道间相关性特征深度挖掘机制,最大限度地描述翻拍过程所引入的颜色通道相关性;还建立了多分辨率深度特征提取网络,提高检测特征对图像分辨率的鲁棒性。本项目的研究成果从基础理论和实际应用两个层面提高翻拍图像检测技术的整体性能,最新的研究成果可以在NTU-ROSE、BJTU-IIS、Dartmouth、ICL-COMMSP等翻拍图像公共数据库上取得超过99%的平均准确率。这些成果为翻拍图像检测技术的进一步研究提供理论基础和技术支撑,促进了该技术在公共安全和智能识别等多个领域的推广应用。
{{i.achievement_title}}
数据更新时间:2023-05-31
内点最大化与冗余点控制的小型无人机遥感图像配准
基于公众情感倾向的主题公园评价研究——以哈尔滨市伏尔加庄园为例
基于全模式全聚焦方法的裂纹超声成像定量检测
基于协同表示的图嵌入鉴别分析在人脸识别中的应用
基于图卷积网络的归纳式微博谣言检测新方法
基于深度特征学习的图像恶劣天气识别与增强技术研究
基于深度特征学习和扩增的图像显著性检测算法研究
基于深度判别特征学习的SAR图像地物分类
基于深度学习与统计特征模型的视频重编码检测技术研究