Discrete Models for High-Level Image Analysis

这个项目开发了用于自动图像分析的统计模型。所开发的图像模型是通过在模板图像中任意放置大量称为小平面的点来定义的。分面点是分层组织的，并且或者布置在分辨率增加的网格上，或者使用更复杂的标准来确定，例如，它们可以位于图像的尺度空间中的局部极值。对象形状是通过连接树状结构中的面，然后使用图像模板中树的形状来建模来自类的观察图像中的面树的预期形状来建立的。通常，3万到4万个面被定位在二维图像模板中，而四分之一到150万个面被放置在三维图像集(例如，磁共振图像或正电子发射断层扫描图像)中。这些模型适用于广泛的成像方式，并且只需要为所调查的一组图像提供模板图像或原型图像。在给定类的图像模板的情况下，可以在很少或没有操作员干预的情况下应用派生模型，以获得对可用图像数据的自动分割、对象识别和体积分析。

9803365Valen E. JohnsonThis project develops statistical models for automatic image analysis. The image models developed are defined by arbitrarily placing a large number of points, called facets, in a template image. The facet points are organized hierarchically, and are arranged either on grids of increasing resolution or determined using more sophisticated criteria, e.g., they may be located at local extrema in the scale-space of the image. Object shapes are established by connecting facets in a tree-like structure, and then using the shape of the tree in the image template to model expected shapes of facet trees in observed images from the class. Typically, thirty to forty thousand facets are positioned within a two-dimensional image template, whereas one-quarter to one-half million facets are placed within three-dimensional image sets (e.g., magnetic resonance images or positron emission tomography images).Models generated under this proposal facilitate computationally efficient high-level analysis of generic images. These models are applicable to a wide range of image modalities, and require only that a template image or prototype image be available for the group of images under investigation. Given an image template for the class, derived models can be applied with little or no operator intervention to obtain automatic segmentation, object identification, and volumetric analysis for available image data.