Content-based image retrieval

计算机视觉研究人员的目标是建立一个通用的计算机视觉系统，能够以与人类相当的方式“看到”。这种进步对社会的影响可以与通用计算机带来的数字革命相媲美。不幸的是，事实证明这项任务极具挑战性。很大程度上的困难在于创建计算机程序的能力，这些程序可以分析数字图像，以检测和识别在未知视觉背景下设置的人造和自然物体。这项任务很困难，因为物体可能出现在图像中的许多空间位置和尺度上，图像背景实际上可以是任何东西，并且在构成特定类别的元素时通常存在很大的视觉变化。本提案解决了计算机视觉中两个相关但不同的问题：特定目标识别；2. 通用对象类检测。特定对象识别利用单个对象特征的先验知识（训练数据），以简化确定图像中人造或自然对象的确切身份的任务。一个例子是识别一个特定的人的脸。相比之下，通用对象类别检测系统的目标是(i)查找和（ii）确定图像中对象的类别身份或类别。一个例子是在毕业班的照片中找到所有的面孔。在这种情况下，需要图像中的空间定位，以及物体的姿势（2D或3D）和大小。这个提议的长期目标是解决过去被称为图像理解或场景分析的问题。也就是说，创建一个能够理解场景的计算机程序，并通过扩展，描述和解释它。从更实际的角度来看，考虑了两种长期情景。一个是开发一个程序，可以自动在网络上搜索特定对象或对象类的图像。第二是创造一种能够与人类环境认真互动的智能视觉机器。也许前者比后者更为现实。

Context of the project within the fieldThe goal of researchers in computer vision is to build a general-purpose computer vision system that can "see" in a fashion comparable to humans. The impact of such an advance on society stands to rival the digital revolution brought about by the general-purpose computer. Unfortunately, this task has proven to be extremely challenging. Much of the difficulty revolves around the ability to create computer programs that can analyze digital images to both detect and identify man-made and natural objects that are set against an unknown visual background. The task is difficult since objects may appear at many spatial locations and scales in the image, the image background can be virtually anything, and there is often a very large visual variation in what constitutes the elements of a particular class.This proposal addresses two related, but different problems in computer vision: 1. Specific object recognition; 2. Generic object class detection. Specific object recognition makes use of a priori knowledge (training data) of a single object's features in order to simplify the task of determining the exact identity of a man-made or natural object in an image. An example is identifying a particular person's face. By contrast, the goal of a generic object class detection system is to (i) find and (ii) determine the class identity or category of the objects in an image. An example is finding all faces in a graduating class image. In this case, spatial localization in the image is required, as well as possibly the pose (in 2D or 3D) and size of the object.The long-term objective of this proposal is to solve what used to be called the image understanding or scene analysis problem. That is, create a computer program that understands a scene, and by extension, describe and explain it. From a more practical point of view, two long-term scenarios are contemplated. One is to develop a program that can automatically search the Web for images of specific objects or object classes. The second is the creation of an intelligent visual machine that is capable of seriously interacting with the human environment. Perhaps the first is more realistic than the second.