| Stereology contains a set of tools and procedures that guide the researcher in estimating a quantity of interest that lies in three-dimensional space, from a set of its two-dimensional sections. Typical quantities of interest are: volume, surface area, area, and ratio among these. The precision of these estimators is measured by the variance. Stereology generally is performed manually by counting the number of intersections between regions or features in the image and a low-resolution test system probe superimposed on the image. It is known that those manual methods are tedious, laborious, and often yield high variance due to coarse resolution of the probe. On the other hand, image processing tools can be used to extract features and to quickly estimate the quantity of interest from an image. Note that image processing often suffers due to imperfection of segmentation, but it has the benefit of using a very high-resolution test system---the pixel. In this work, the derivation of the tradeoff between computer-based and manual stereology is explicitly made for the estimate of a volume, and a ratio between two measured volumes. The factors affecting the variance using sampling design are: the spacing interval between slices, the slice thickness, and whether the sampling follows a uniform, a nonuniform, or an adaptive scheme. The inequality of the tradeoff formula is expressed in terms of those factors. From this inequality, the conditions under which one method is preferable to the other become clear. Illustrations are presented as well; stereological principles are applied to clinical data to determine which CT protocol to use in detecting small changes in lung volume. Furthermore, the inequality formula provides a lower bound on the performance that an image processing algorithm must have for it to compete with the existing gold standard: manual stereology methods.; An automatic target recognition (ATR) algorithm developed to recognize seven vehicle types, using two types of sensors (uncooled infrared and visible CCD-Intensity) is presented. The targets are seen in several orientations and at various ranges/sizes. Features that are scale- and orientation-independent, and independent of lighting and meteorological conditions, were extracted from the segmented images. The features describe intensity, shape, and thermal characteristics of the vehicle. Fusion at the feature level was used to determine whether the additional information provided by each sensor can yield better overall performance in classification of these vehicles. The confusion matrix and correct classification rate were indicators used in evaluating the classifiers. The significance of this work lies in the method's ability to use fusion information to classify vehicles accurately even as variations occur in background, range/size, and orientation. In addition, principles of stereology were used to define an adaptive k nearest-neighbor classification method. |
