特征空间的鲁棒性分析：彩色图像分割.pdfVIP

Robust Analysis of Feature Spaces: Color Image Segmentation Dorin Comaniciu Peter Meer Department of Electrical and Computer Engineering Rutgers University, Piscataway, NJ 08855, USA comanici @ meer @ Abstract Nonparametric density estimation [4, Chap. 61 avoids A general techniquefor the recovery of significant im- the use of the normality assumption, however it requires ad- agefeatures is presented. The technique is based on the ditional input information (type of the kernel, number of mean shift algorithm, a simple nonparametric procedure neighbors). This informationmust be provided by the user, for estimating density gradients. Drawbacks of the current and for multimodal distributions it is difficult to guess the methods (including robust clustering)are avoided. Feature optimal setting. Nevertheless, a reliable general technique space of any nature can be processed, and as an example, for feature space analysis can be developed using a simple color image segmentation is discussed. Thesegmentation is nonparametric density estimation algorithm. In this paper completely autonomous, only its class is chosen by the usel: we propose such a technique whose robust behavior is supe- Thus, the same program can produce a high quality edge rior to methods employing robust estimators from statistics. image, orprovide, by extracting all the significant colors, a preprocessorfor content-based query systems. A 512 x 512 2. Requirements for Robustness color image is analyzed in less than 10seconds on a stan- Estimat