Visual Image Interpretation in Humans and Machines

视觉对人类来说是如此基础，这在很大程度上是一个自动化的过程，对我们来说似乎非常容易。这表明，让计算机像人一样看东西应该很容易。事实上，这是一项非常困难的任务，因为生物视觉系统非常复杂，约占人脑的四分之一。人类的视觉既高效又高效。例如，它能够识别大约10,000个不同的对象类别，并可以从单个示例中学习新的类别。这是通过一个只需要20瓦功率、重1.4公斤的系统实现的。在识别能力、学习效率和功耗方面，没有任何计算机系统可以与之匹敌。设计新的计算机视觉方法的方法之一是了解生物视觉系统的工作原理。然而，视觉的复杂性使这一点变得非常困难，一些研究人员将精力集中在理解生物视觉上，而另一些研究人员则寻求独立的解决方案来解决计算机视觉中的特定问题。例如，人类可以阅读车牌，但我们使用的是一个通用的视觉系统，它也可以阅读哥特式文字和笔迹，以及执行许多其他任务。建立一个车牌识别系统，以像人类一样的一般方式阅读字母将是困难的。然而，因为车牌有一定的固定格式(它们总是某种明亮的颜色，字体总是某种样式和大小)，所以构建一个仅仅为了阅读车牌而不是其他东西的计算机视觉系统，是一项简单得多的任务。事实证明，对于计算机视觉来说，有些任务并不简单，在这些任务中，理解生物视觉可能对未来的成功至关重要。一个例子是匹配两个曲面的外观。假设您想要使人造石头看起来与建筑物中的真实石头一模一样。为了得到正确的配方，你不仅需要知道原始石头的物理性质(这可能无法完全匹配)，而且还需要知道人类视觉系统可能如何感知石头。然后，你可以选择一个配方，它可能不完全模仿石头，但对人类来说，它看起来就像真正的石头。此外，如果你知道视觉系统如何处理表面的颜色和纹理，你就可以建立一个计算机化的工具，可以自动预测食谱。另一个感兴趣的领域是计算机图形学。让计算机图形看起来令人信服的一种方法是对你试图表示的事物的物理模型进行精确建模。然而，这种渲染方法通常非常耗时且计算代价高昂。由于人类视觉系统并不能看到物体中的每一个细节，因此利用人类视觉系统将如何处理每个场景的知识来更快、更有效地绘制图形通常是可能的。由于那些致力于生物视觉的研究人员往往来自生物学和心理学背景，而那些研究计算机视觉的人来自计算机科学和工程背景，这两组研究人员之间经常存在理解上的差距，这使得他们很难在上面概述的问题上合作。该网络的目的是使这些研究人员在物质上和科学上更紧密地联系在一起，以便他们能够在最有可能成功的具有挑战性的问题上确定和合作。我们将通过一系列非正式会议以及资助初级科学家和博士生在不同学科的另一个实验室工作来实现这一点。

The sense of vision is so fundamental to humans that it is a largely automated process which appears to us as extremely easy. This would suggest that it should be easy to make a computer see like a human. In fact this is a very difficult task because the biological visual system is very complex; occupying about one quarter of the human brain. Human vision is both highly effective and efficient. For example it is capable of identifying around 10,000 different object categories and can learn new categories from single examples. This in achieved with a system requiring just 20 watts of power and weighing 1.4kg. No computer system can match this performance for recognition ability, learning efficiency and power consumption.One way to devise new computer vision methods is to understand how biological visual systems work. However, the complexity of vision has made this very difficult and some researchers have concentrated their efforts on understanding biological vision while others have sought independent solutions to specific problems in computer vision. For example, humans can read car number plates but we do so using a general purpose visual system that can also read gothic script and handwriting as well as performing a host of other tasks. Building a number plate recognition system to read letters in the same general way that humans do would be difficult. However, because number plates have a certain fixed format (they are always a certain, bright, colour, and the font is always a certain style and size) building a computer vision system just to read number plates, and nothing else, is a much simpler task. There are some tasks that have not proved simple for computer vision and where understanding biological vision is likely to be essential to future success. One example is matching the appearance of two surfaces. Suppose you wanted to make artificial stone to look exactly like the real stones in a building. To get the recipe just right you would have to know not just the physical properties of the original stone (which probably cannot be matched exactly) but also how the human vision system is likely to perceive the stone. You can then pick a recipe that may not mimic the stone exactly but which will look just like the real stone to humans. Moreover, if you know how the visual system processes the colours and textures of surfaces you can build a computerised tool that can predict recipes automatically. Another area of interest is computer graphics. One way to make computer graphics look convincing is to exactly model the physics of the thing you are trying to represent. However, such rendering methods are often very time consuming and computationally expensive. Because the human visual system does not see every detail in an object it is often possible to render graphics much more quickly and effectively using perceptual rendering techniques that exploit knowledge of how the human visual system will process each scene.Because those researchers working on biological vision tend to be from Biology and Psychology backgrounds and those who research computer vision from Computer Science and Engineering backgrounds, there is often a gap in understanding between the two groups of researchers which makes it hard for them to work together on problems such as those outlined above. The aim of this Network is to bring such researchers closer together, both physically and scientifically, so that they can identify and work together on the challenging problems where success is most likely. We will achieve this by a series of away day style meetings and conferences and by funding junior scientists and PhD students to spend time working in another lab from a different discipline.