文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > 人文社科 > 文学研究 > Comparing and Evaluating Interest Points
侵权投诉

Comparing and Evaluating Interest Points

Kitchen and Rosenfeld 13] present an interest point detector which uses the curvature of planar curves. They look for curvature maxima on isophotes of the signal. However, due to image noise an isophote can have an important curvature without corresponding to an interest point, for example occurring on a quasi-uniform region. Therefore, the curvature is multiplied 2+ 2;2 xx y yy x xy x y . Dreschler by the gradient magnitude of the image. Their measure is K= 2+ 2 x y and Nagel 8] developed an approach equivalent to the one of Kitchen. Harris 10] has improved the approach of Moravec by calculating a matrix which is related to the auto-correlation function. Compared to the approach of Moravec it is no longer necessary to use discrete directions and discrete shifts. This matrix averages the rst derivatives of the signal on a window:"# 2 2 y2 I I I; x 2+2 exp I I I2 I I I I I I I I I x x y x y y

The eigenvalues of this matrix are the principal curvatures of the auto-correlation funct

ion. If these two curvatures are high, an interest point is present. Cottier 5] has changed the implementation of the Harris detector. To improve the localization of the detected points, he only applies the detector to the contours of the images and successively uses two di erent window sizes to increase precision of detection. Forstner 9] proposes a feature extraction method based on local statistics of the image function. Image pixels are classi ed into categories - region, contour and interest point - by using the auto-correlation function. The use of statistics allows a blind estimation of signal-dependent noise variance and thus an automatic selection of thresholds. The method developed by Heitger et al. 11] was inspired by experiments on the biological visual system. Their interest point measure seems to be representative of special cells of the human vision system. They rst extract 1D directional characteristics by convolving the image with orientation-selective Gabor lters. In order to obtain 2D characteristics, they compute the rst and second derivatives for all of the 1D characteristics. The detector of Reisfeld et al. 20] uses the concept of symmetry. They compute a symmetry map which shows a\symmetry strength" for each pixel. This symmetry is computed locally by looking at the magnitude and the direction of the derivatives of neighbouring points. Points with high symmetry are selected as interest points.

2.1.3 Methods based on template tting Methods based on template tting are used to obtain sub-pixel accuracy. They t the image signal to a parametric model of a speci c type of interest point, for example corner or vertex. There are therefore not applicable in a general context. Rohr 21] ts the signal to a template in the case of junctions of several lines. The template is a binary junction model which is convolved with a Gaussian. In the case of a corner the parameters of the model are the angle of the symmetry axis, the angle of the corner, the greylevels, the position of the point and the blur. These parameters are adjusted such that the template is closest to the observed signal. Corners obtained by this method are very precise. However, the quality of the approximation depends on the initial position estimation. Rohr uses the intersection of line segments to determine the initial parameters of the model. 5

上一页第5页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email