基于Gauss滤波的图像可视化检索(Ⅱ)——全局匹配

基于多尺度Gauss滤波的图像特征提取 ,能够在较大的尺度和视角变化范围内 ,检索出具有相似外观特征的图像。在讨论多尺度Gauss滤波输出用于图像的局部相似性匹配检索的基础上 ,进一步分析了如何将多尺度Gauss滤波方法用于全局相似性匹配检索。首先由图像在不同尺度下的Gauss滤波输出 ,构造出图像的曲率和相位直方图 ,得到图像的多尺度直方图表示 ;然后查询图像和数据库图像的归一化的曲率和相位直方图再进行匹配。归一化的直方图因消除了图像中点的位置信息 ,因此是图像变换如 :旋转、扭曲、拉伸下的不变量。该方法能用到基于图像外观特征的全局可视化检索中 ,并能取得很好的效果

Image Retrieval in Multimedia Databases Based on Gauss-Filtering (Part II)——Global Similarity

Gauss-filtering based on multi-scale feature extraction of images has been proved an effective method. It is invariable to the object's position, image-plane orientation, scale and view change, so it can retrieve images from database by the similarity of their visual appearance. In this paper, we first discuss local similarity image retrieval using multi-scale Gauss-derivative filters. Then we generalize the idea to a wide range of image retrieval by global similarity. We use the Gauss filtering output to construct curvature and phase histogram of images. By comparing the normalized curvature and phase histogram between the template image and database images, we can retrieve image in database based on their global similarity. Our simulation shows that the result is satisfying and the parameters can be tuned by user's preference. Now, we decide to integrate the local similarity and global similarity method of Gauss derivative filters into a single framework for user's convenience.