基于视觉显著内容的图像半脆弱自恢复水印算法研究

To trade off the robustness of semi fragile self-recovery watermarking algorithm and the quality of recovery, this research is to deeply study the semi fragile self-recovery algorithm of image watermarking based on visual salient contents. First, the effective representation of watermarking information is proposed. Afterwards, we propose a watermark generation method based on visual salient contents and compressive sensing, which can not only reduce the watermark amount and but also improve the fault tolerance of the generated watermark. Second, a block-wise image multi-bit watermark embedding and extraction solution is proposed. For this solution, we propose a randomly-group quantization embedding method for increasing the robustness of the algorithm to the tolerable distortion. Moreover, we propose a Bit Error Rate (BER) model via a minimum-measure-based watermark extractor, which chooses the fine-tuned distortion compensation factor to accurately extract the watermarks when the image content integrity is kept. Finally, the solutions of the tamper localization, identification, image recovery are proposed. Namely, we propose a method based on a double authentication and clustering method to increase the accuracy of tamper localization and employ a SVM classifier to identify the tamper type, and accordingly select the suited recovery strategy for reconstructing the salient image content. This research aims to improve the image recovery quality while retaining the algorithm robustness. This research offers new ideas for the semi-fragile self-recovery watermarking and highlights its potential of practical values and applications.

针对半脆弱自恢复水印算法鲁棒性及图像恢复质量之间存在的矛盾，本课题拟对基于视觉显著内容的图像半脆弱自恢复水印算法进行深入研究。首先，研究水印信息的有效表示。拟提出基于视觉显著内容及压缩感知的水印生成方法，降低水印量，同时增加水印的容错性。其次，研究基于图像块的多比特水印嵌入及提取策略。拟提出随机分组量化的水印嵌入方法，增加算法对于合理失真的鲁棒性，并基于最小测度水印提取器，建立误码率模型，选取合适的失真补偿因子，在未遭受恶意篡改的前提下，准确提取水印位；最后，研究篡改定位、篡改类型判别及图像恢复策略。拟提出基于双向认证及聚类分割的篡改定位方法，提高篡改定位精度，并使用SVM分类器对破坏类型进行判别，针对性地选取恢复策略，重构图像显著内容。课题研究将在保证算法鲁棒性的基础上，提高图像恢复质量。同时，该课题也将为半脆弱自恢复水印研究提供新思路，并进一步提高其实际应用价值，拓宽应用前景。

针对半脆弱自恢复水印算法鲁棒性及图像恢复质量之间存在的矛盾，本课题对基于视觉显著内容的图像半脆弱自恢复水印算法进行了深入研究，包括: 1) 视觉显著内容提取，分别提出了“基于纹理差异的视觉显著性模型”，“基于纹理及上下文分析的视觉显著性模型”，“基于局部统计与整体分析的视觉显著性模型”，“基于稀疏优化的视觉显著性模型”等，为显著性模型建立及显著内容提取提供了新方法；2) 水印鲁棒嵌入和提取，提出了“显著差异特征量化的鲁棒水印嵌入算法”，“显著幅值差异特征量化的鲁棒水印嵌入算法”，“最大系数修改及阈值分割的鲁棒水印算法”，为半脆弱自恢复的鲁棒水印嵌入与提取提供了新思路；3) 篡改检测及自恢复，提出了“基于小波分组量化与双认证策略的半脆弱自恢复水印算法”；最后将研究成果应用到了人脸保护中，提出了“基于半脆弱自恢复水印的人脸图像保护算法”。 共发表论文13篇，包括国际SCI源期刊论文5篇，EI源期刊1篇，国际会议论文4篇，中文核心期刊论文3篇；申请发明专利2项；待出版学术专著1部；获河南省自然科学优秀学术论文一等奖2篇。该课题为半脆弱自恢复水印研究提供了新思路，并进一步提高了其实际应用价值，拓宽了应用前景。