Object and Concept Recognition for Content-Based Image Retrieval

这项研究的目标是开发必要的方法来自动识别数字图像中的通用对象和概念类(如建筑物、汽车、船只和树木)，以便显著改进基于内容的图像检索过程，该过程主要依赖于低级颜色和纹理特征来匹配查询和数据库图像。该方法有三个主要方面：(1)设计新的高层图像特征，包括将低层特征和关系特征组合在一起的簇特征，这些特征捕捉它们之间的空间关系；(2)开发一个统一的表示，可以以学习系统可以使用的形式表达大量的低层和高层特征；(3)通过使用分层的多分类器方法来自动化对象和概念类的识别器的开发。由此产生的技术正在几个不同的大型图像数据库上进行评估，其中包括商业数据库，其图像被分成广泛的类别，以及一个提供每幅图像中物体列表的地面真实数据库。这项工作的结果将是一种新的通用对象和概念识别范例，可以立即应用于大型图像数据库的自动或半自动索引。该方法将有助于弥合图像检索系统用户的高级需求和通常从图像中提取的低级特征之间的差距。我们开发的通用目标类识别算法将开启目标识别研究的新时代，离开几何领域，进入概念领域。

The goal of this research is to develop the necessary methodology for automated recognition of generic object and concept classes (such as buildings, cars, boats, and trees) in digital images in order to substantially improve the process of content-based image retrieval, which has relied mainly on low-level color and texture features for matching queries to database images. The approach has three major aspects: (1) to design new high-level image features including cluster features that group together lower-level features and relationship features that capture spatial relationships among them; (2) to develop a unified representation that can express a large variety of both low- and high-level features in a form that can be used by learning systems; and (3) to automate the development of recognizers for object and concept classes through the use of a hierarchical, multiple classifier methodology. The resulting techniques are being evaluated on several different large image databases, including commercial databases whose images are grouped into broad classes and a ground-truth database that provides a list of the objects in each image. The results of this work will be a new generic object and concept recognition paradigm that can immediately be applied to automated or semi-automated indexing of large image databases. The methodology will help to bridge the gap between the high-level needs of users of image retrieval systems and the low-level features typically extracted from an image. The generic object class recognition algorithms we develop will begin a new era of object recognition research, leaving the geometric domain and entering the conceptual domain.