ARRA: Identifying Objects Within Scenes: Combining Context and Features in Visual Object Recognition

ARRA：识别场景中的对象：在视觉对象识别中结合上下文和特征

• 批准号: 0958615
• 负责人: Elan Barenholtz
• 金额: $ 19.66万
• 依托单位: Florida Atlantic University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2010
• 资助国家: 美国
• 起止时间: 2010-09-15 至 2013-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0958615&HistoricalAwards=false
• 关键词: ARRA; Identifying; Objects; Within; Scenes

This award is funded under the American Recovery and Reinvestment Act of 2009 (Public Law 111-5).Despite nearly half a century of research, the human ability to recognize objects visually remains a largely unsolved puzzle. Previous research on object recognition has primarily considered cases in which the target is viewed in isolation. However, the visual system can use contextual information -- such as the presence of other objects in the scene or knowledge about the kind of environment in which the object is found -- to determine the identity of an object as well. The contribution of this kind of information is especially clear when the image of the object itself is insufficient on its own. For example, a small yellow patch might be identified as a partially obscured banana in the context of a fruit bowl or as a leaf in the context of a tree. In an NSF-funded research project, Dr. Elan Barenholtz at Florida Atlantic University will use behavioral and computational techniques to examine two central questions regarding the role of context in object recognition: 1) How do people acquire knowledge about the relations between objects and their contextual scenes (for example, the likelihood of specific objects appearing in a certain type of context)? 2) How is this knowledge put to work in recognizing objects whose images have been degraded and cannot be recognized on their own? This research will employ two experimental methodologies: The first will use computer-generated artificial scenes in which participants must first learn the object/context relations from scratch and later use this knowledge in a recognition task. The second technique will test object recognition abilities in photographs of real world environments, including pictures of participants' own homes or workplaces. In this case, subjects will have knowledge about the expected object/context relations, based on their long-term experience, particularly when the environment is highly familiar to them. Human performance in these tasks will be assessed using statistical methods to assess the contribution of contextual information in object recognition.Understanding human visual object recognition holds great promise for brain science -- as much as a third of the human cortex is thought to be devoted to visual processing. Such understanding is also important for designing artificial vision systems, which carry an enormous array of potential applications. However, previous theoretical techniques, which focused on specialized algorithms for extracting 3-dimensional structure from individual objects, have proven largely unsuccessful. Dr. Barenholtz's research represents a strong departure from earlier approaches, as it assumes that visual recognition relies on inferential strategies that draw on an individual's broad knowledge about the world and his or her experience with specific environments. This approach treats vision as relying on similar tools as other cognitive processes, such as inference and decision-making, suggesting that there may be a great deal of previously unexplored common ground across these different disciplines. By putting the "cognition" back into "recognition," this research has the potential to contribute to some long awaited breakthroughs in the field of visual recognition.

该奖项由2009年《美国复苏和再投资法案》(Public Law 111-5)资助。尽管进行了近半个世纪的研究，但人类视觉识别物体的能力在很大程度上仍然是一个悬而未决的谜团。以前关于目标识别的研究主要考虑了孤立地观察目标的情况。然而，视觉系统可以使用上下文信息--例如场景中其他对象的存在或关于发现该对象的环境类型的知识--来确定对象的身份。当物体本身的图像本身不充分时，这种信息的贡献尤其明显。例如，在水果碗中，一个小的黄色斑块可能被识别为部分遮挡的香蕉，或者在树的上下文中，可能被识别为树叶。在NSF资助的一个研究项目中，佛罗里达大西洋大学的Elan Barenholtz博士将使用行为和计算技术来研究关于背景在对象识别中的作用的两个核心问题：1)人们如何获得关于对象与其上下文场景之间的关系的知识(例如，特定对象出现在特定类型的上下文中的可能性)？2)如何利用这些知识来识别图像已退化且无法自行识别的对象？这项研究将采用两种实验方法：第一种方法是使用计算机生成的人工场景，参与者必须首先从头开始学习对象/上下文关系，然后在识别任务中使用这些知识。第二项技术将在真实世界环境的照片中测试物体识别能力，包括参与者自己的家或工作场所的照片。在这种情况下，受试者将根据他们的长期经验，特别是在他们非常熟悉环境的情况下，了解预期的对象/背景关系。人类在这些任务中的表现将通过统计方法进行评估，以评估背景信息在物体识别中的贡献。理解人类视觉物体识别对脑科学来说是很有希望的--据认为，多达三分之一的人类大脑皮层致力于视觉处理。这样的理解对设计人工视觉系统也很重要，人工视觉系统承载着大量潜在的应用。然而，以前的理论技术侧重于从单个对象中提取三维结构的专门算法，已被证明在很大程度上是不成功的。巴伦霍尔茨博士的研究与以前的方法大相径庭，因为它假设视觉识别依赖于推理策略，这些策略依赖于一个人对世界的广泛知识以及他或她对特定环境的经验。这种方法将视觉视为依赖于与其他认知过程(如推理和决策)类似的工具，这表明在这些不同的学科中可能存在大量以前未被探索的共同点。通过将“认知”放回到“识别”中，这项研究有可能为视觉识别领域期待已久的一些突破做出贡献。