快速图像前景分割和流形Matting

OPT OpenIR > 研究生部

	快速图像前景分割和流形Matting
	刘康
学位类型	硕士
导师	陶大程
	2016-05
学位授予单位	中国科学院大学
学位授予地点	北京
关键词	计算机视觉图像前景分割超像素流形学习片对齐
摘要	视觉信息在人类接受的信息中约占83 %，尤其是在大数据时代，作为视觉信息的主要载体，图像、视频的有效处理和分析非常有价值。而图像分割是辅助视觉信息智能化处理、图像内容理解的研究热点和基础性问题，它在建立底层特征与高层语义联系中有承接作用。经过几十年的研究探索，研究人员提出来大量的图像分割方法，但是图像分割仍然普遍存在时间复杂度高、分割精度有待提高、目标语义难以自适应提取等问题。其中，图像前景分割是图像分割的一个特例，针对图像前景分割中的效率和精度问题，本文做了两方面研究。针对大多数图像前景分割的时效性低、需要人工标记三分图（Trimaps）的问题，本文提出了一种基于超像素GrabCut 和图像Matting 技术的交互式快速前景分割算法。特别地，首先应用超像素分割算法对图像进行预处理，然后利用GrabCut 分割超像素图像并获得粗略超像素图像的Mask。接着，本文又提出一种快速自适应的Trimaps 生成算法（FATs）。最后，通过应用图像Matting 技术交互式迭代处理得到最终改善的前景Mask。实验在BSDS500 和Alphamatting 数据库上定性和定量地证明了提出算法的有效性，在视觉表观，MSE, SAD 和E-Time 等综合指标上都优于对比算法。针对图像Matting 问题，本文提出了基于流形学习思想的图像Matting 框架，命名为片对齐流形Matting（PAMM）。首先提出了图像颜色空间的局部部分模型，然后将颜色局部流形子空间共享到Alpha 子空间，并通过全局对齐优化得到逼近Alpha 空间的重构误差公式。最后，应用高效Nesterov 算法来迭代求解该优化问题。作为流形Matting 框架的一个应用实例，本文也提出了一些新的流形Matting 算法，如ISOMAP Matting 和 Boosting ISOMAP Matting (BIM)。我们把Alphamatting 作为实验数据库，因为该数据库提供了用于Matting 对比的公开的公用Trimaps。实验结果验证了流形Matting 框架的有效性。另外，因为ISOMAP 能够很好地处理非线性数据分布并保持像素类别的判别性，因此实验从定性和定量两方面证明了ISOMAP Matting 和BIM 的最好表现。
其他摘要	It is revealed that the vision information plays about 83% role of the whole information which is received by human. Especially in the age of big data, it is worthwhile to do analysis of images and videos. Therefore, image segmentation is a hot topic and basic problem in the field of understanding and intelligentized processing of vision semantics. Moreover, image segmentation make a connection between the low level features and the high level semantics. Over decades, although an ocean of segmentation algorithms are proposed, there are still some universal problems, such as time complexity, segmentation accuracy and self-adaptive image semantics. In particular, the foreground segmentation is a research brach of image segmetation. Aimming at the efficiency and accuracy problems of foreground segmentation, in this thesis there are twofold researches as follows. Aimming for the problem that most of foreground segmentation algorithms are timeconsuming, and the trimaps used in the matting step are labelled manually. In this thesis, we propose a fast interactive foreground extraction method based on the superpixel GrabCut and image matting. Specifically, we frst extract superpixels from a given image and apply GrabCut on them to obtain a raw mask. Due to that the resulting mask border is hard and toothing, we further propose fast and adaptive trimaps (FATs), and construct a FATs-based Shared matting for computing a refined mask. Finally, by interactive processing, we can obtain thefinal foreground. Experimental results on the BSDS500 and alphamatting datasets demonstrate that our proposed method is superior to four representative methods whether in the vision view or in the three evaluation criteria, MSE, SAD and E-Time. Aimming at the existing problem of matting, in this thesis a manifold based matting framework named Patch Alignment Manifold Matting (PAMM) is firstly proposed. In particular, we first propose a part model of color space in the local image patch. Furthermore, we perform whole alignment optimization for approximating the alpha results by using subspace reconstructing error. Finally, we utilize an efficient Nesterov algorithm to solve the optimization problem. As an application of the framework, some new manifold learning matting algorithms, such as named ISOMAP Matting and its derived Boosting ISOMAP Matting (BIM), are also proposed. Alphamatting dataset is selected as the benchmark for the competitive matting algorithms in the experiment. The experimental results show the effectiveness of the manifold matting framework. In addition, the ISOMAP Matting and BIM can deal with the nonlinear data distribution and better preserve discriminability of pixel classes. Furthermore, the good performance of the ISOMAP Matting and BIM is demonstrated, both qualitatively and quantitatively.
学科领域	计算机应用其他学科（含图像处理）
语种	中文
文献类型	学位论文
条目标识符	http://ir.opt.ac.cn/handle/181661/27965
专题	研究生部
作者单位	中国科学院西安光学精密机械研究所
推荐引用方式 GB/T 7714	刘康. 快速图像前景分割和流形Matting[D]. 北京. 中国科学院大学,2016.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
快速图像前景分割和流形Matting.p（8152KB）	学位论文		限制开放	CC BY-NC-SA	请求全文