ITR: Learning and Measuring Perceptual Similarity

图像检索多年来一直是一个活跃的研究领域，但有两个基本问题仍然没有得到很大的解决：1)如何最好地学习用户的主观查询概念；2)如何准确地度量感知相似度。第一个问题涉及提出查询概念的完备性，例如，如何提出诸如“动物”、“大教堂”或“飞机”之类的查询。第二个问题涉及搜索精度，即给定一个已学习的查询概念，如何找到与该概念匹配的所有图像。为了解决这两个基本问题，并确保我们的解决方案具有可扩展性，本项目有四个具体目标。首先，我们计划开发新的主动学习算法，在不受时间和样本限制的情况下，快速学习用户的主观查询概念(思想和意图)。其次，我们将设计半自动的图像标注和标注精化方法，为图像赋予语义标签，以支持多通道查询-概念学习和信息检索。第三，我们将设计感知距离函数来提高视觉搜索的准确性。例如，一旦了解了诸如‘敌舰’这样的查询概念，我们希望在监视数据库中找到每一个匹配的对象，而不是遗漏任何对象。最后，我们计划使用由不同机构(包括IBM研究中心和旧金山美术馆)的同事提供的实验数据，对开发的学习算法进行验证研究。

Image retrieval has been an active research area for many years, buttwo fundamental problems remain largely unsolved: 1) How best tolearn users' subjective query concepts, and 2) How to measure perceptualsimilarity with significant accuracy. The first problem concerns thecompleteness of formulating a query concept, e.g., how to formulate aquery such as ``animals,'' ``cathedrals,'' or ``aircraft.'' The secondproblem concerns search accuracy, i.e., given a learned query concept,how to find all images that match that concept.To tackle these two fundamental problems and to ensure that oursolutions are scalable, this project has four specific targets.First, we plan to develop novel active learning algorithms thatquickly learn users' subjective query concepts (thoughts and intents)despite time and sample constraints. Second, we will designsemi-automatic image annotation and annotation refinement methodsfor assigning semantic labels to images in order to support multimodalityquery-concept learning and information retrieval. Third, we willdevise perceptual distance functions for improving accuracy of visualsearches. For instance, once a query concept such as ``enemy vessels''is learned, we want to find every matching object in the surveillancedatabase, not missing any. Finally, we plan to conduct validationstudies} on developed learning algorithms, using experimental dataprovided by colleagues at various institutions (including IBM researchcenters and Fine Arts Museums of San Francisco).