数字水印中的几何不变特征点提取算法研究-智能与分布计算实验室

数字水印中的几何不变特征点提取算法研究

姓名	陶茂垣
论文答辩日期	2007.01.26
论文提交日期	2007.02.02
论文级别	硕士
中文题名	数字水印中的几何不变特征点提取算法研究
英文题名	Research on Geometric Invariant Feature Points Extraction Algorithm in Watermarking
导师1	卢正鼎
导师2
中文关键词	特征点提取;数字水印;曲率尺度空间;边缘检测;角点检测
英文关键词	Feature Extraction;Digital Watermarking;Scale Space;Edge Detection;Corner Detection
中文文摘	图像几何不变特征点提取对第二代抗几何攻击数字水印的设计与实现有着极其重要的意义，它为局部水印嵌入提供参考点，关系到嵌入水印的鲁棒性。几何不变特征点提取还对模式识别，人脸器官的识别定位，基于内容的图像视频检索技术都有着非常重要的意义。基于曲率尺度空间（curvature scale space）的抗几何变换图像特征点是一种基于轮廓的特征点提取方法。在提取图像特征点之前，首先需要提取图像的轮廓信息，然后在不同的尺度下观察轮廓，找到曲率最大的点来作为特征点，然而由于提取图像轮廓计算量很大，所以算法整体效率不高。Harris-Laplace角点检测方法是一种直接基于灰度图像多尺度抗几何攻击角点提取算法，但需要对每个尺度都进行特征点提取，计算也比较复杂。将Harris-Laplace角点检测方法进行改进，把直接分析图像局部灰度值的角点提取方法与图象尺度空间的思想相结合，并兼顾多尺度的不同权值，则既可以保证角点抵抗一般几何攻击的鲁棒性，又减少计算复杂度，根据此思路提出了加权平均Harris-Laplace角点检测方法来提取特征点的方法，为每个尺度指定一个权值，取多尺度图像取加权平均值作为特征点提取的依据，这样就只需要在加权平均的特征增强图像上提取一次特征点，减少了计算量。通过大尺度观察图像，可以得到图像的粗糙画面；而从小尺度观察，能够检测到图像的细节特征。根据这个思路又提出了迭代逼近Harris特征点提取算法，该算法从大尺度开始提取特征点，逐步降低尺度，迭代逼近特征点的真实位置，不但减少了噪声对图像的影响，同时也减少了运算量。在以上研究的基础上，设计并且实现了一个基于图像尺度空间的特征点提取算法评价平台，实验表明，以上的两种算法既能很好的保证特征点提取的质量，同时计算时间比传统特征点提取算法明显减少。
英文文摘	Geometrical invariant feature extraction, which can provide reference points, has been a key factor to second generation image or video watermarking technology. It also plays a very important role in pattern recognition, face and organ recognition and image content searching. Feature points extraction based on curvature scale space needs to detect the edge of the image before detecting feature points. After the edge is detected, the image is observed under different scales and the points attaining the local maxima in curvature are selected as feature points that are invariant to geometrical transforms. However, the algorithm is not so effective to meet the real-time requirement because the edge detection is costs too much time. Harris-Laplace corner detector is a multi-scale gray level image corner detector, but the computation is very complicated because each level of scaled image must be calculated to extract feature points. By improving Harris-Laplace algorithm, a novel algorithm is presented, which combines scale space theory and Harris detector and gives every scale a weight value to mark the importance of every weight. Using the weighted average Harris response, the feature points are stable to some geometrical transforms such as cropping and scaling while the computation is relatively smaller. Observing the image through large scale can aqquire coarse content of the images while observing it through small scale can get fine content of images. According to this hypothesis, iterative convergent Harris detector is another corner detector based on scale space which first extracts feature points on the coarse scale and traces them to locate the exact position while the scale becomes fine. In this way, the algorithm has better performance against noise while reducing the whole computation. On the base of the above-mentioned research, an experiment platform of benchmarking feature points extractions is designed and established. According to the experiment, the proposed algorithms can not only guarantee the quality of the extracted feature points but also reduce the computation time compared to conventional algorithms.