• 遥感科学与应用技术 •

### 结合ASIFT和归一化的抗仿射变换遥感影像盲水印算法

1. 1.兰州交通大学测绘与地理信息学院,兰州 730070
2.地理国情监测技术应用国家地方联合工程研究中心,兰州 730070
3.甘肃省地理国情监测工程实验室,兰州 730070
• 收稿日期:2021-01-15 修回日期:2021-04-10 出版日期:2021-10-25 发布日期:2021-12-25
• 通讯作者: * 张黎明（1975— ）,男,甘肃天水人,博士,教授,主要从事地理信息安全,空间数据安全,版权保护,地图数据数字水印的理论与方法等研究。E-mail: zhang_lm@163.com
• 作者简介:秦如贞（1997— ）,女,山西吕梁人,硕士生,主要从事遥感影像版权保护研究。E-mail: 1043020927@qq.com
• 基金资助:
甘肃高等学校产业支撑引导项目项目(2019C-04);国家自然科学基金项目(41761080);兰州交通大学优秀平台支持(201806)

### A Blind Watermarking Algorithm for Remote Sensing Image based on Anti-Affine Transforma-tion Combining ASIFT and Normalization

QIN Ruzhen1,2,3(), ZHANG Liming1,2,3,*(), WU Tingchen1, LI Yu1,2,3, WANG Hao1,2,3

1. 1. Faculty of Geomatics, Lanzhou Jiaotong University, Lanzhou 730070, China
2. National-Local Joint Engineering Research Center of Technologies and Applications for National Geographic State Monitoring, Lanzhou 730070, China
3. Gansu Provincial Engineering Laboratory for National Geographic State Monitoring, Lanzhou 730070, China
• Received:2021-01-15 Revised:2021-04-10 Online:2021-10-25 Published:2021-12-25
• Supported by:
Industrial Support and Guidance Projects of Colleges and Universities in Gansu(2019C-04);National Natural Science Foundation of China(41761080);Lanzhou Jiaotong University Excellent Platform(201806)

Abstract:

In the process of using remote sensing images, the synchronization between watermark and remote sensing image is disrupted after affine transformation is performed on the watermarked image. Although the watermark is not removed after affine transformation, the size and relative position of the image are changed, which leads to the problem that watermark cannot be detected correctly. It is significant to improve the robustness of digital watermarking. This paper proposes a blind watermarking algorithm for remote sensing image, which can effectively resist affine transformation. Firstly, the affine invariant feature points of the image are extracted by ASIFT algorithm. The stable, evenly distributed, appropriate feature points are selected according to the size of the feature scale and the Euclidean distance between the feature points to construct the corresponding square feature regions. Secondly, take a feature region as an example, the affine invariant moments of the feature region are calculated to normalize the feature region. With the invariant centroid of the normalized feature region as the center, the sub region is extracted as watermark embedding region. The low-frequency information of the watermark embedding region is obtained by two-level wavelet transform. The watermark is embedded into the low-frequency information using the quantization embedding rule to complete the watermark embedding. Finally, inverse wavelet transform of low-frequency information is performed. The difference image of the characteristic region before and after embedding is denormalized. The denormalized difference image is superimposed on the original image characteristic region to complete the watermark embedding. In this paper, a GF-2 remote sensing image of 3000 × 3000 pixels was used as the carrier image, while the binary image with copyright information was used as the watermark. This paper measured the invisibility of the watermark by calculating the magnitude of the peak signal-to-noise ratio. In the robustness testing experiments, the watermarked image was subjected to affine transformation, noise adding attack, and filtering attack. Then, the algorithm of this paper was used for watermark extraction. The watermark with the highest normalized correlation value extracted was selected for robustness evaluation. The experiment results show that, after affine transformation, including rotation, translation, and scaling, the watermark information was accurately extracted from the watermarked remote sensing image through the algorithm. The algorithm is robust to noise, filtering, cropping, and other conventional watermarking attacks. The watermark has good imperceptibility. ASIFT is a blind watermarking algorithm. The watermark detection process does not need the original remote sensing image. Our algorithm has strong practicability.