CAREER: Image Metrics for Computer Graphics

这个项目将研究在计算机图形学的一系列重要应用中使用感知动机的图像度量。 计算机图形学中的各种各样的任务假设两个图像之间的外观相似性的某种程度。 这些任务包括细节层次生成、光栅字体设计、制图元素放置、纹理合成和形状变换。 在这些工作中使用的指标往往不能很好地反映人类对给定任务的视觉感知。 该项目将重新访问计算机图形学中的几个领域，并将基于视觉的指标应用于许多应用程序，以产生更好的图像。 用基于感知的图像相似性度量取代现有应用中使用的技术，有可能在广泛的重要应用中改善结果。 可以从使用图像指标中受益的应用包括： - 飞行员和驾驶员训练模拟器。 - 机械零件的计算机辅助设计。 - 自动化地图制作 - 光栅字体创建。 - 大型建筑物内部的交互式漫游，用于建筑审查 - “模型”器官与医学患者的MRI或CT数据之间的比较 - 矢量和张量场可视化应用，如天气预报和机械和建筑结构的应力/应变模拟。 该项目将研究使用图像处理和计算机视觉技术来帮助定义特定任务的图像指标。 这些度量可能包含的一些属性是：对边缘强度和方向的敏感性，在许多尺度上的频率响应，图像内强度和颜色的相对而非绝对测量，以及噪声对感知的掩蔽效应。 一旦定义了度量，将使用优化过程来最小化生成的图像或几何图像与目标对象之间的差异。 优化技术的选择将取决于应用和图像度量的形式。 计算上昂贵的优化方法可以接受离线任务。 在速度至关重要的情况下，可以通过在执行优化例程时观察优化例程来建议优化策略。 该项目的教育目标是向本科生和研究生介绍计算机科学研究，使学术界以外的社区接触计算机图形学的重要应用，并将该项目开发的程序，库和数据集传播给世界各地的其他研究人员。 学生将通过正在开发的计算机图形课程和研究小组内的工作参与研究。 ***

This project will investigate the use of perceptually-motivated image metrics across a range of important applications in computer graphics. A wide variety of tasks in computer graphics assume some measure of similarity in appearance between two images. Such tasks include level-of-detail generation, raster font design, cartographic element placement, texture synthesis and shape transformation. Too often the metrics used in such work are poor at reflecting what a human would visually perceive for the given task. This project will re-visit several areas within computer graphics and will apply visually-based metrics to a number of applications with the goal of producing better images. Replacing the techniques that are used in existing applications with perceptually-based measures of image similarity has the potential to improve the results in a wide range of important applications. Applications that can benefit from the use of image metrics include: - Simulators for pilot and driver training. - Computer aided design of mechanical parts. - Automated map-making. - Raster font creation. - Interactive walkthroughs of large building interiors for architectural review - Comparison between "model" organs and the MRI or CT data of a medical patient - Vector and tensor field visualization for applications such as weather prediction and stress/strain simulations of mechanical and architectural structures. This project will investigate using techniques from image processing and computer vision to help define image metrics for particular tasks. Some of the properties that these metrics may incorporate are: sensitivity to edge strength and orientation, frequency responses at a number of scales, relative rather than absolute measures of intensity and color within an image, and the masking effects that noise has on perception. Once a metric is defined, an optimization procedure will be used to minimize the differences between the generated image or geom etry and the target object. The choice of optimization technique will depend on the application and on the form of the image metric. Computationally expensive optimization methods may be acceptable for off-line tasks. Where speed is essential, heuristics may be suggested by observations of the optimization routine while it is executing. The educational goals of this project are to introduce both undergraduate and graduate students to computer science research, to expose the community outside of academia to important applications of computer graphics, and to disseminate the programs, libraries and datasets that are developed by this project to other researchers worldwide. Students will be involved in research both through the computer graphics classes that are being developed and by working within a research group. ***