An Integrated Method for Simultaneous Recognition and Segmentation of Deformable Objects

本研究主要旨在从自然图像中稳健地定位近似已知形状的可变形结构/物体。定位和识别图像中潜在物体结构的问题在许多图像分析和计算机视觉应用中都很重要，包括机器人视觉、模式识别和生物医学图像处理。能够可靠地执行这些基本任务的算法是各种系统的核心，这些系统将允许人类更有效地与图像数据进行交互，以实现可视化、分析、控制的目的，或者在图像数据库系统的情况下，检索这些应用领域内的特定任务信息。这种结构的鲁棒识别和测量并不总是可以实现使用单一的分析技术，依赖于单一的图像派生的信息来源。当处理在具有不同图像内容的不同条件下获得的广泛图像时，尤其如此。本研究项目中的方法利用了两个不同的图像衍生信息来源：1)灰度梯度和ii)强度或纹理元素等的均匀性，以及基于模型的信息（即物体的近似形状）。此外，该方法结合了两种不同的处理方法，以整合上述信息源：a)基于区域的方法，主要基于同质性；b)同时利用梯度（图像派生）和形状（基于模型）属性的边界方法。将使用博弈论框架，其中，不像在计算机视觉中广泛用作集成方法的全局目标方法，保留了潜在目标的模块化。然后将集成问题构建为一系列耦合且共存的目标，其中一个模块的输出依赖于其他模块先前的输出。这项工作将建立在能够在二维图像中定位结构的初步工作的基础上。它还将概念扩展到分析三维图像数据。将对该方法的适用性和局限性进行理论和实验研究。

This research is primarily aimed at robustly locating deformable structures/objects of approximately known shape from natural images. The problem of locating and recognizing underlying object structures in an image is of importance in many image analysis and computer vision applications including robot vision, pattern recognition and biomedical image processing. Algorithms that can reliably perform these basis tasks are at the core of a variety of systems that will permit humans to more effectively interact with pictorial data for the purposes of visualization, analysis, control or, as in the case of image database systems, retrieval of task-specific information within these applications areas. The robust identification and measurement of such structure is not always achievable using a single analysis technique that depends on a single image derived source of information. This is especially true when one is dealing with a wide range of images obtained under different conditions having different image content. The approach in this research project utilizes two different sources of image-derived information: 1.) gray-level gradients and ii.) homogeneity of intensity or texture elements, etc., as well as model-based information (i.e approximate object shape). Furthermore, two different processing methods are combined in this approach, in order to integrate the above-mentioned information sources: a) regionbased methods, which are primarily based on homogeneity properties and b.) boundary methods which capitalize on both gradient (imagederived) and shape (model based) properties. A game theoretic framework will be used, where, unlike the global objective approach which is widely used as an integration method in computer vision, the modularity of the underlying objectives are retained. The integration problem is then framed as a family of coupled and coex isting objectives whereby the output of one module depends upon the previous outputs of the other modules. This effort will build on initial work which has been able to locate structure in two dimensional images. It will also extend the concept to analyzing three dimensional image data. Both theoretical and experimental investigations will be carried out regarding the applicability and the limits of the approach.