CAREER: Learning Models for Scalable Content-Based Image Retrieval

职业：可扩展的基于内容的图像检索的学习模型

• 批准号: 0952943
• 负责人: Lorenzo Torresani
• 金额: $ 49.5万
• 依托单位: Dartmouth College
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2010
• 资助国家: 美国
• 起止时间: 2010-04-01 至 2017-03-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0952943&HistoricalAwards=false
• 关键词: CAREER; Learning; Models; Scalable; Content

This project addresses the design of machine learning algorithms enabling content-based image retrieval in Web-scale collections of photos. This research formulates image retrieval as a binary classification problem: decide which database images are the "same" as the user-provided photo. Efficiency and scalability to large collections are achieved by constraining the classifiers to be models supported by traditional text-search engines, which perform real-time search in databases of several billion documents. In order to implement search based on high-level notions of similarity, the research team develops methods to automatically localize the most content-relevant regions in the input photo and to extract from them semantically powerful classifiers combining appearance cues with robust geometric constraints. The algorithms learn from user-provided labels indicating the presence but not the location of similar visual content, thus requiring a minimal amount of human supervision. This research investigates also how this advanced form of similar-image search can be used to organize personal photos, provide semantic annotations, and support content-based clustering of pictures. Furthermore, this work provides technical advances in a wide range of computer vision problems including object detection, visual saliency, and content-based clustering of photos. Moreover, the research team is collecting an unprecedentedly large image data set to evaluate the developed image retrieval system and to be available to the community. Research is naturally integrated with education and outreach by means of related courses and out-of-classroom activities aimed at attracting students to this field and at encouraging interdisciplinary collaborations.

该项目解决了机器学习算法的设计，使基于内容的图像检索能够在web规模的照片集合中实现。本研究将图像检索描述为一个二元分类问题：决定哪些数据库图像与用户提供的照片“相同”。通过将分类器约束为传统文本搜索引擎支持的模型，可以实现对大型集合的效率和可伸缩性，传统文本搜索引擎在数十亿个文档的数据库中执行实时搜索。为了实现基于高级相似性概念的搜索，研究团队开发了自动定位输入照片中与内容最相关的区域的方法，并从中提取语义强大的分类器，将外观线索与鲁棒几何约束相结合。算法从用户提供的标签中学习，这些标签表明了类似视觉内容的存在，而不是位置，因此需要最少的人工监督。本研究还探讨了如何使用这种高级形式的类似图像搜索来组织个人照片，提供语义注释，并支持基于内容的图片聚类。此外，这项工作为广泛的计算机视觉问题提供了技术进步，包括目标检测、视觉显著性和基于内容的照片聚类。此外，研究小组正在收集前所未有的大量图像数据集，以评估开发的图像检索系统，并向社会提供。通过相关课程和旨在吸引学生进入这一领域并鼓励跨学科合作的课外活动，研究自然与教育和推广相结合。