Parametric Analysis and Transfer of Pictorial Style

有些绘画风格差异很大，不需要仔细观察就能区分出来，例如，点彩画派与文艺复兴;或者大师的照片与随意的快照。 研究人员探索使用定量测量的绘画风格，导致参数表征，捕捉粗粒度的绘画风格。 这项研究在计算机图形学和图像分析中有几个应用，对社会产生了广泛的技术影响。它可以嵌入到图像分类和检索工具中。 对于数码摄影，它可以将大师摄影师的风格转移到随意的快照中，从而提高质量。 此外，它将导致直接作用于风格方面的新颖和直观的图像处理工具。 最后，这样的统计估计可能有助于在未来了解是什么使图像逼真，并导致过滤器，提高现实主义。 这个研究项目是伴随着一个跨学科的课程“艺术和科学的描绘”，并通过一个跨学科的图像统计讲习班。本研究研究的边缘和联合统计的导向多尺度分解，如可控金字塔。 研究人员开发了一种新的递归金字塔分解，可以捕获图像上“纹理”的空间变化。他们还探索了边缘特征的统计数据，并设计了新的非线性边缘保持分解，以防止在风格转换过程中在强边缘处出现光晕。 最后，本研究探索了一种基于色类命名概念的色彩风格表征新方法。 颜色的纯度被定义为它与颜色原型（例如纯蓝色）的距离，并且利用该纯度的统计数据来评估和增强颜色鲜艳度。 通过分类任务（监督学习）和风格转移来评估风格表征的成功：在目标图像中强制执行源图像或图像集的相关统计。 目标图像的视觉修改允许独立于内容地评估哪些风格方面被统计捕获。 该研究建立在视觉感知，图像分析和计算机图形学之间的协同作用。

Some pictorial styles differ dramatically and can be distinguished without scrutinizing the picture, for example, Pointillist vs. Renaissance; or a photograph by a master vs. a casual snapshot. The investigators explore pictorial style using quantitative measurements, leading to a parametric characterization that captures coarse-grain pictorial style. The research has several applications in computer graphics and image analysis, with broad technological impact on society. It can be embedded in image classification and retrieval tools. For digital photography, it can transfer the style of master photographers to casual snapshots, leading to important quality enhancement. In addition, it will lead to novel and intuitive image manipulation tools that act directly on stylistic aspects. Finally, such statistical estimators might help in the future to understand what makes images photorealistic and lead to filters that increase realism. This research project is accompanied by an interdisciplinary course on the "art and science of depiction," and by an inter-disciplinary workshop on image statistics. The research studies marginal and joint statistics of oriented multiscale decompositions such as steerable pyramids. The investigators develop a new recursive pyramidal decomposition that captures the spatial variations of "texturedness" over an image. They also explore the statistics of edge features, and they devise new non-linear edge-preserving decompositions to prevent haloing artifacts at strong edges during style transfer. Finally, the research explores a new approach to color style characterization based on the notion of naming color category. The purity of a color is defined as its distance to a color prototype such as pure blue, and statistics of this purity are leveraged to assess and enhance color vividness. Success of the style characterization is evaluated by classification tasks (supervised learning) and by style transfer: the relevant statistics of a source image or set of images are enforced in a destination image. The visual modification of the destination image permits the assessment of which stylistic aspects are captured by the statistics, independently of content. The research builds on the synergy between visual perception, image analysis and computer graphics.