基于压缩感知和变换域的数字图像隐私保护技术研究

Compressive sensing (CS), as a new theory for signal acquisition, has been widely applied in the fields of information theory, image processing, wireless communication, optics and pattern recognition etc. In the recent years, the researchers attempt to apply the compressive sensing theory to design the image encryption and information hiding algorithms. In the current CS-based privacy protection schemes, there exist some fundamental defects such as high computational complexity, low efficiency, large key volume, inability to resist the chosen-plaintext attack and so forth. To overcome these drawbacks, this project will conduct the following research work: (1) Based on CS theory, the union dictionary/the sparse over-complete dictionary construction method, the atomic grouping/classification approach, the clustering algorithms and the image processing techniques, the new image sparse representation schemes with excellent performance will be developed to achieve good performances of image compression and reconstruction. The influence of the sparse base, the image block size and the sampling rate on the image reconstruction quality and execution efficiency will be also investigated. (2) Based on the aforementioned studies, the fast, secure and robust fusion-compression-encryption and compression-encryption schemes for the multi-focus images and multiple images will be designed by using CS, chaos, cryptography, selection and encryption for Region of Interest, the image fusion technology and the image processing methods. And their performance will also be analyzed. (3) CS, chaotic encryption, the information hiding techniques, the transform methods in frequency domain and Optical image processing schemes will be adopted to develop the fast, secure and robust large capacity information hiding algorithms in encrypted images together with the above obtained image compression-encryption results. The aims of this project are outlined as follows: (1) simultaneously realize image compression and encryption; (2) implement the large capacity information hiding in ciphered images based on CS; (3) further develop the CS theory and extend its application scope. The proposed schemes can be applied to the privacy protection of multimedia information in the network and cloud environments.

压缩感知理论是一种全新的信号采集理论。近年来，人们尝试将其应用于设计数字图像加密和信息隐藏方案，但存在着计算复杂度高、效率低、安全性不高等问题。对此，本项目将开展如下研究：（1）基于压缩感知理论、联合字典/过完备原子库构造、原子分组/分类、聚类算法和图像处理技术，设计性能优良的图像信号稀疏表示新方法，并寻找稀疏基、图像分块大小、采样率等对图像重构效果和效率的影响关系；（2）在此基础上，利用压缩感知、兴趣区域加密、图像融合和图像变换技术等，设计快速、安全、鲁棒的（多聚焦、多幅）图像（融合-）压缩-加密算法，并分析算法性能；（3）基于前述研究结果，并结合使用信息隐藏技术、频域变换技术和光学图像处理技术等，设计快速、安全、鲁棒的图像密文域大容量信息隐藏方案。本项目目的在于实现数字图像压缩和加密的同步实施，实现图像密文域大容量信息隐藏，发展压缩感知理论并应用于网络和云环境下多媒体信息的隐私保护。

不同于经典采样定理，压缩感知理论是一种全新的信号采集理论。近年来，国内外学者提出了许多基于压缩感知的数字图像隐私保护方案，但存在着计算复杂度高、效率低、安全性差等缺陷。针对以上问题，我们对数字图像压缩加密和密文域大容量信息隐藏进行了深入研究。首先，我们设计了基于布尔卷积神经网络与池化技术、图像分块离散小波信号、图像动态自适应滤波技术和KSVD与PCA的信号稀疏表示新方法，研究了稀疏基、图像分块大小、采样率等因素对图像重构效果和效率的影响。其次，利用压缩感知、深度学习、图像融合和图像变换技术等，我们设计了一系列快速、安全、鲁棒的（多幅）图像（融合-）压缩-加密算法：基于二维压缩感知、拉丁方变换和小波基，设计了一种双彩色图像压缩-加密算法；基于混合多目标粒子群优化、块压缩感知和Hessenberg分解，设计了一种有效的视觉上有意义的多彩色图像融合-压缩-加密方案；基于压缩感知、GAN、CNN和混沌系统，提出了一种鲁棒图像压缩加密算法；设计了一种有限计算精度环境下的基于三维布尔卷积神经网络的高安全图像加密方案；提出了一种云环境下基于遗传算法的保留缩略图加密技术，使加密后的图像与原始图像可以生成完全一致的缩略图，而缩略图中保留的粗略明文信息可以提供部分视觉可用性，方便数据管理和图像检索；基于图像滤波、DNA加密、拉丁方置乱、改进的遗传算法和矩阵半张量积等，提出了至少五种不同的彩色图像加密算法。最后，基于二维压缩感知和信息隐藏技术，设计了一种双彩色图像密文域大容量信息隐藏方案；基于奇异值分解、DWT、Hachimoji DNA加密和Julia集，设计了两种快速、安全、鲁棒的双彩色图像密文域大容量可逆信息隐藏方案；设计了一种用于图像缩略图保留加密的多层离散小波变换信息自嵌入方法。发表高水平学术论文20篇，申请国家发明专利7件，并尝试将研究成果应用于网络和云环境下数字图像的隐私保护。