Efficient Description, Modeling, and Recognition of Natural Imagery via a Local Basis Library

9973032研究人员开发自适应，分层，计算效率高的算法，用于描述，建模和识别自然图像，通过紧密结合计算谐波分析技术和来自生物视觉系统研究的洞察力。 具体的研究目标是：1）研究有效表征给定图像类别的基本图像特征，并利用这些特征开发计算效率高的随机图像模型用于仿真：2）开发分层算法，通过集成基本特征来提取高级复杂特征，以构建更好的随机模型;以及3）研究这种随机模型用于识别和分类不同图像类别的有效性。 在整个研究过程中，研究者充分利用所谓的本地基础库作为基本工具来提取基本特征，这些基本特征是图像相对于该库中包含的基础的展开系数（或这些系数的非线性函数）。 该局部基库由局部基字典（如小波包、局部余弦、局部傅立叶和刷波）的集合组成，其基元素在空间和空间频率上都是局部化的。 每个字典又包含大量的正交或双正交基，并配备了一个快速的数值算法来选择一个最好的基础，适合于一个给定的任务和一组给定的图像，通过优化特定任务的标准。 在目标1中，研究者通过显式地减少它们之间的冗余和统计依赖性来从库中专门选择局部特征，并研究稀疏性和统计独立性对于图像表示和描述的重要性。 然后，他通过假设它们之间相互独立或成对依赖来构建简单的随机模型，分析这些特征的有效性。 在目标2中，研究人员将目标1中研究的局部特征集成到这些局部特征上，通过将Karhunen-Loeve变换的本地化版本直接应用于这些局部特征来提取高级复杂特征，并将这些特征纳入随机模型中。 在目标3中，研究者检查结合线性判别分析的本地化版本和目标1和目标2获得的模型的有效性，用于提取关键的判别局部features.Results的项目将是有用的，对于那些经常工作的基于图像的诊断，如放射科医生和地质学家。 放射科医生使用非常微妙的特征，乳房X光片，以确定是否病人有良性或恶性肿瘤。 勘探地质学家使用从实际岩石样本到地震图像的多种尺度的图像来确定特定的地下层是否含有石油或天然气。 尽管这些专家拥有训练有素的眼睛，并通过训练和经验在头脑中积累了大量的图像数据库，但他们的诊断规则通常是定性的。 研究者认为，上述特征提取工具将帮助这些专家将他们的诊断方法从定性制度转向定量。 事实上，给定特定类别的图像（例如，表示恶性肿瘤的乳房X线照片的集合），上述算法可以允许这些专家建立指定图像的关键特征的统计依赖性的随机模型，并从这些模型创建新的样本图像。 这种随意创建新的模拟图像的能力对于评估特征和图像的可变性是必不可少的。 如果专家能够对这些模拟图像给出正确的诊断，那么这意味着他们头脑中的隐式诊断规则被编码为基于这些特征提取工具计算的特征的显式随机模型。 这可能导致分布式数字图像诊断系统的发展，其中成像设备和诊断引擎/知识库位于远程。 人们可以很容易地想象这种系统在医学、地球物理勘探和其他需要图像诊断的领域中的有用性。

9973032The investigator develops adaptive, hierarchical, and computationally efficient algorithms for description, modeling, and recognition of natural imagery by tightly combining computational harmonic analysis techniques and the insight derived from the study of biological vision systems. The specific study aims are to: 1) Investigate the elementary image features that efficiently characterize a given image class, and develop computationally efficient stochastic image models using these features for simulation; 2) Develop hierarchical algorithms to extract high-level complex features by integrating the elementary features for building better stochastic models; and 3) Investigate the effectiveness of such stochastic models for recognizing and classifying different image classes. Throughout this study the investigator fully utilizes the so-called local basis library as a basic tool to extract elementary features that are the expansion coefficients (or nonlinear functions of such coefficients) of an image with respect to the bases contained in this library. This local basis library consists of a collection of the local basis dictionaries (such as wavelet packets, local cosine, local Fourier, and brushlets) whose basis elements are localized in both space and spatial frequency. Each dictionary in turn contains a huge number of orthogonal or biorthogonal bases, and is equipped with a fast numerical algorithm to select a best basis tailored to a given task and to a given set of images by optimizing task-specific criterion. In Aim 1, the investigator specifically selects the local features from the library by explicitly reducing the redundancy and statistical dependence among them, and investigate the importance of the sparsity and statistical independence for image representation and description. Then, he analyzes the effectiveness of such features by constructing simple stochastic models by assuming mutual independence or pairwise dependence among them. In Aim 2, the investigator integrates local features studied in Aim 1 to extract high-level complex features by applying the localized versions of the Karhunen-Loeve transform directly on those local features, and incorporates such features into the stochastic models. In Aim 3, the investigator examines the effectiveness of combining the localized versions of the linear discriminant analysis and the models obtained by Aims 1 and 2, for extracting the key discriminant local features.Results of the project will be useful for those who routinely work on image-based diagnostics such as radiologists and geologists. Radiologists use very subtle features of mammograms to determine whether the patients have benign or malignant tumors. Exploration geologists use images of multiple scales ranging from actual rock samples to seismic images to determine whether a specific subsurface layer contains oil or gas. Although these experts have highly trained eyes and have accumulated large databases of images in their minds through training and experience, their diagnostic rules are often of qualitative nature. The investigator believes that the above feature extraction tools will help these experts move their diagnostic methods from a qualitative regime to a quantitative one. In fact, given a specific class of images (e.g., a collection of mammograms representing malignant tumors), the above algorithms may allow these experts to build stochastic models that specify the statistical dependency of critical features of the images and to create new sample images from these models. This ability to create new simulated images at will is indispensable for assessing the variability of the features and images. If an expert can give a correct diagnosis for these simulated images, then this implies that the implicit diagnostic rules in their minds are encoded as an explicit stochastic model built on features computed by these feature extraction tools. This may lead to the development of a distributed digital image diagnostic system, where imaging devices and the diagnostic engines/knowledge bases are remotely located. One can easily imagine the usefulness of such systems in medicine, geophysical exploration, and other fields requiring image diagnostics.