基于多尺度边缘感知的图像平滑和分层编辑研究

Edge-aware image smoothing and hierarchical editing has recently emerged as a valuable tool and received considerable attention from researchers. It breaks out the limitation of traditional camera by producing images with more visual conflict and attraction. However, the traditional edge-aware image smoothing and hierarchical editing methods cannot effectively manipulate edges　with different scales, and usually lead to losing textures, fusing salient edges and preserving noise in image processing. The research of this project aims at studying the method to identify and manipulate multi-scale edges in image smoothing and hierarchical editing. The main research includes the following three parts: (1) Study the method of multi-scale edge aware image smoothing. By separately editing edges with different scales, we will design a new smoothing method which can preserve textures, enhance edges and filter noises in image smoothing. This method can be used to filter image while preserving its multi-scale edges. (2) Build an edge-adapting iterative image smoothing model which can adaptively adjust the enhancements of edges with different scales in each smoothing. We will use this model to propose a scheme to iteratively smooth an image while preserving its edges. This scheme can be used to produce artistic and stylized images. (3) Based on the new scheme of iterative image smoothing, we will build a new method of image hierarchical editing which supports adjusting distorted edges, edge aware image decomposition. This hierarchical editing method can be used to implement tone mapping and detail enhancement by avoiding blurring edges and introducing gradient reversal. We will spread the achievement of this project on the digital life platform such as providing super-resolution images without noise or artistic images for digital television.

边缘感知的图像平滑和分层编辑可突破传统相机的限制，生成更多视觉冲突、更有吸引力的图像，具有重要应用价值，是当前的研究热点。然而传统主流的方法无法合理控制不同尺度的边缘，常丢失纹理细节、模糊显著边缘、残留噪声。本项目致力于研究感知和操控多尺度边缘的图像平滑和分层编辑方法，主要研究：（1）多尺度边缘感知的图像平滑方法。通过分离和编辑不同尺度的边缘，设计出保留纹理、增强显著边缘、消除噪声的平滑方法，实现多尺度边缘保持的图像滤波；（2）建立边缘自适应的迭代平滑模型，自动调整每次平滑时各尺度边缘被锐化的程度，提供边缘保持的迭代平滑方案，可用于生成艺术化风格的图像；（3）基于新设计的迭代平滑方案，构建调整了失真边缘的、基于边缘感知图像分解的图像分层编辑方法，以提供无边缘模糊、无梯度反转的色调映射和细节增强方法。研究成果将在数字家庭平台上推广应用，如为数字电视提供更高清的画面或艺术化风格的图片。

本项目研究基于多尺度边缘感知的图像滤波和分层编辑方法。首先，结合迭代非对称采样、局部线性模型、边缘矫正算法和双边滤波，实现了大尺度边缘和纹理细节的分离研究；为提高多尺度边缘分离的效率和有效性，还进一步利用局部全变分模型和拉普拉斯金字塔模型，实现了大尺度边缘、小尺度边缘、纹理、细节的分离；基于以上研究的边缘感知和分离方法，我们建立了能够保持大尺度边缘、增强小尺度边缘、抑制纹理细节的图像平滑算法。然后，基于多尺度边缘的感知和识别，根据具体应用需求，展开图像分层编辑研究，具体包括：实现了能够保持大尺度边缘和压缩全局亮度范围的高动态图像映射算法、构造了可丰富纹理细节且不导致边缘翻转的细节增强模型、设计了基于多层次纹理感知的图像去雾方法、提出了图像结构保持的颜色转换方法、构造出基于大尺度边缘信息的图像重定位编辑模型。最后，还通过设计边缘密度算法，将边缘感知的图像处理引入视频内容分析，提出了边缘信息引导的目标跟踪方法。总体上，课题执行状况良好，研究任务按计划圆满完成，共发表共发表国内外期刊和国际会议论文15篇，其中SCI索引5篇，EI索引7篇；申请相关专利和软件著作权3项，其中软件著作权已授权；参与培养博士1名，培养研究生4名；指导学生参加比赛获奖2次（省部级一等奖1次，三等奖1次）；完成任务计划书中规定的考核指标。