CIF: Small: RUI: Fixation-Driven Contour Integration of Natural Images for Early Visual Processing

一个可靠的自动图像理解系统可以对许多应用产生巨大的影响，包括图像搜索、视频监控、自动驾驶汽车和机器人等等。然而，与文本理解和语音识别相比，该技术的进展缓慢。这个问题可以归因于缺乏将图像分解成一组有意义的组件的方法，类似于文本和语音处理中的单词。本研究将开发一种算法，有效地将数字图像分割成这些有意义的组件。为了实现这一目标，研究人员将重点放在颜色和亮度的不连续点上，这些不连续点通常被称为“边缘”，并研究将它们分组和描绘图像中物体的算法方法。本科生将积极参与涉及计算机科学、数学、统计学、心理学和生物学的跨学科研究。PI还将开发一门跨学科课程，整合认知、神经和计算机科学。由此产生的源代码、软件工具、数据和可视化材料将向公众提供，以促进STEM教育。更具体地说，研究集中在研究人员最近开发的两项创新：线性滤波器矩阵的半群平滑，以及越来越复杂的图的连续划分。前者是一种仿射交换线性算子，可以有效地提取高曲率点，同时具有抗混叠的鲁棒性。它将用于平滑、分割和表征轮廓碎片。后者从简单形状到复杂形状依次在焦点周围生成一组闭合轮廓。此外，该研究还研究了新的上下文敏感感知显著性度量，该度量基于边缘周围的环境来量化边缘的重要性。

A reliable system for automated image understanding can have immense impacts on many applications including image search, video surveillance, autonomous vehicles, and robotics, to name a few. However, the progress of the technology has been slow compared to text understanding and speech recognition. The problem can be attributed to lacking ways of breaking an image into a set of meaningful components analogous to words in text and speech processing. This research will develop an algorithm to partition a digital image into such meaningful components efficiently and effectively. To reach the goal, the investigators focus on discontinuities in color and brightness often called "edges" and study algorithmic ways to group them and delineate objects found in the image. Undergraduate students will actively participate in interdisciplinary research involving computer science, mathematics, statistics, psychology and biology. The PI will also develop an interdisciplinary course that integrates cognitive, neuro, and computer sciences. The resulting source code, software tools, data, and visual materials will be made publicly available to promote STEM education.More specifically, the study centers on two recent innovations developed by the investigators: semi-group smoothing with a matrix of linear filters, and successive partitioning of a graph with increasing complexity. The former is an affine commutative linear operator that is shown to be effective in extracting high-curvature points while robust against aliasing. It will be used to smoothen, partition, and characterize contour fragments. The latter generates a set of closed contours surrounding a focal point successively from a simple shape to more complex ones. Furthermore, the study investigates new context sensitive perceptual saliency metrics that quantify the importance of edges based on their surroundings.