Illuminating Colour Constancy: from Physics to Photography

在日常生活中，我们依赖于代表现实世界的彩色图像，从关键个人事件的照片到可能购买的照片。总体而言，这些都是对真实情况的糟糕近似。我们的目标是更好地理解我们在现实世界中如何感知颜色，以及如何用图像重建这种感知。这些目标的核心是颜色恒定，这是一种基本现象，即使在照明颜色发生巨大变化的情况下也能保持对象颜色的稳定-无论是在蓝光下还是在黄钨光下，我们都认为苹果是红色的。忠实地记录变化的光信号的相机传感器并不自然地具有颜色稳定性。但数码相机通常配备了特殊的色彩平衡模块，以应对光线的变化，它们产生的照片可能会进行进一步处理，以消除色彩投射。在计算机视觉中，为了使机器能够使用颜色作为可靠的提示--例如，在瓷砖等制成品的自动分级中--这种“颜色校正”算法是必要的。人类视觉和计算机视觉通常是相互孤立地研究的：第一个目的是理解为什么颜色看起来像它们对人类所做的那样，另一个目的是使它们尽可能地对机器有用，无论它们是如何出现的。这两个目标通常不是完全相同的，因为人类和计算机的颜色恒定都不是完美的。为了将颜色恒定从人类连接到机器，我们将进行一系列创新的实验。首先，我们将系统地研究光源的同色异谱作用。同色异谱是所有图像复制的基础：两种颜色光谱截然不同的刺激可以诱导出相同的颜色感知。与白天一张白纸的平坦光谱反射比相比，电视上发出的白光具有高度尖峰的光谱。然而，照亮白纸时看起来相同的光源有时会使其他表面改变外观。当我们购买在人造商店灯光下看起来很好，但当我们把它们带到户外时就不那么令人满意的衣服时，我们就会经历这种现象。我们将使用一种新的可调节的光谱照明器来量化真实灯光下的真实场景的这种效果，我们可以使用它来生成任何光谱。我们的第二个创新是利用新上市的高动态范围(HDR)显示器。与真实世界形成鲜明对比的是，场景中最亮的点可能是最暗点的100000倍，大多数显示器难以产生1000：1的动态范围，打印的照片至多是100：1。然而，我们知道色彩感知取决于场景的整体动态范围。新的HDR显示器可以输出100000：1的对比度，我们将使用它们在实验室条件下测量稳定性，但具有真实世界的亮度。要使彩色照片与我们对现实世界的感知相匹配，我们面临的第三个挑战是颜色记忆的不准确性。通常，当我们观看一张照片时，我们没有真实的东西可以与之比较，而是必须从记忆中回忆起原始的场景。那么，我们记忆的缺陷可能会玷污我们的判断力。众所周知，我们对天空、草和皮肤等熟悉物体的记忆颜色往往是“过饱和的”--人们记忆中的草可能比实际更绿，天空可能比实际更蓝。因此，当我们通过询问人们喜欢哪种图像来测试颜色校正算法时，我们可能会发现他们并不喜欢最准确地再现原始场景的图像，而是与他们不完美的记忆相匹配。我们将量化这些记忆和偏好的影响。最后，我们的研究将在所有阶段考虑如何通过数学模型预测可测量的颜色知觉。最终，我们将设计算法，像我们一样自动识别颜色，从而创造出更好的照片和更有用的视觉机器。

In daily life, we depend on colour images which represent the real world, from photographs of key personal events to pictures of possible purchases. In general, these are poor approximations of the real thing. Our aim is to understand better how we perceive colours in the real world, and how to recreate that perception with images. Central to these aims is colour constancy, a fundamental phenomenon which keeps object colours stable even under large changes in the colour of the illumination - we see an apple as red whether it is under bluish daylight or yellowish tungsten light. Camera sensors, which faithfully record the changing light signals, do not naturally possess colour constancy. But digital cameras are often equipped with special colour balancing modules to cope with changes in lighting, and the photographs they produce may be further processed to remove colour casts. In computer vision, such 'color correction' algorithms are necessary to enable machines to use colour as a reliable cue - for example, in automated grading of manufactured goods such as tiles. Human vision and computer vision are typically studied in isolation from each other: the first aims to understand why colours appear as they do to humans, and the other to make them as useful as possible to machines, regardless of how they appear. These two goals are generally not identical, because neither human nor computer colour constancy is perfect.To bridge colour constancy from humans to machines we will perform an innovative set of experiments. First, we will systematically study illuminant metamerism. Metamerism is what makes all image reproduction work: two stimuli with vastly different colour spectra can induce the same colour percept. The light invoking a white percept on a TV has a highly spiky spectrum compared to the flat spectral reflectance of a piece of white paper in daylight. Yet, illuminants which look the same when shining on white paper can sometimes make other surfaces change appearance. We experience this phenomenon when we buy clothes which look good under the artificial shop lights but less satisfactory when we take them outdoors. We will quantify this effect for real scenes under real lights using a new 'tuneable' spectral illuminator with which we can generate any light spectrum. Our second innovation is to make use of newly available High-Dynamic-Range (HDR) displays. In contradistinction to the real world where the brightest point in the scene may be a 100000 times as bright as the darkest point, most displays struggle to produce a dynamic range of even 1000:1 and printed photographs are at most 100:1. Yet we know that colour perception depends on the overall dynamic range of the scene. The new HDR displays can output contrast ratios of 100000:1 and we will use them to measure constancy in lab conditions but with real world brightnesses. A third challenge that we face in making colour photographs match our perception of the real world is the inaccuracy of colour memory. Typically, when we view a photograph, we do not have the real thing to compare it with, but must recall the original scene from memory. The imperfections of our memory then may taint our judgment. It is well known that our memory colours for familiar objects such as sky, grass, and skin tend to be 'over-saturated' -- grass may be remembered as greener and the sky as bluer than they actually are. Thus, when we test colour correction algorithms by asking people which image they prefer, we might find that they do not prefer the one that most accurately reproduces the original scene, but instead matches their imperfect memory. We will quantify these effects of memory and preference. Finally, our research will, at all stages, consider how measured percepts of colour might be predicted by mathematical models. Ultimately, we will design algorithms to automatically see colours as we do, making for better photographs and more useful vision machines.