基于分层图结构化稀疏低秩表示的目标联合分割方法研究

项目名称： 基于分层图结构化稀疏低秩表示的目标联合分割方法研究

项目编号： No.61502244

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 计算机科学学科

项目作者： 葛琦

作者单位： 南京邮电大学

项目金额： 20万元

中文摘要： 联合分割针对一组图像上一系列具有相似特征的共同目标进行识别和分割。当前大部分联合分割方法采用局部化和全局化相关性分离的建模方式，由于不同图像上背景和目标的复杂多变、扰动等因素影响，这种建模方式并不准确且影响时效。本项目将局部化和全局化相关性结合，提出一套新的基于分层图和图结构化稀疏与低秩表示的显著性目标联合分割框架。首先基于联合显著性检测出共同目标的潜在区域；然后基于点、块、形状三种特征构建基于低层视觉特征的分层3D马尔科夫图模型和基于形状的3D马尔科夫图模型，获得初步分割的结果；最后，研究结合稀疏低秩表示和分层3D图模型，构建基于图结构化稀疏低秩表示的分割传播模型，修复初步分割的结果提高分割精度。本项目基于视觉感知机制，将局部和全局相关性系统的融合在的分割模型中，并拓展了稀疏低秩表示模型的应用层面，所提供的方法也是后续图像和视频的各项处理任务的理论依据和技术支撑。

中文关键词： 联合分割；分层图；结构化稀疏；低秩

英文摘要： Cosegmentation recognizes and segments the similar common objects from different images. Recently, most methods build cosegmentation model based on local correlation firstly and then based on global correlation. However, as the results of variation of background and foreground, this modeling strategy by most cosegmentation methods is not accurate and may lead to low efficiency. This project integrates these two correlations together, propose a novel cosegmentation framework of hierarchy 3D MRF graph and graph structured sparse low-rank represented saliency model. First, the latent regions where the common objects is on are detected by joint saliency detection; Second, the low-level feature based hierarchy 3D MRF graph model is built by features from point, local patch, and another 3D MRF graph is built by shape feature. By optimizing these two models in order, we obtain the primary cosegmentations. At last, we integrate the sparse and low-rank represented model with hierarchy 3D MRF graph, and build graph structured sparse and low-rank represented model to propagate segmentations for repairing the bad segmentations from primary segmentation. This project according to human visual perception mechanism, fuses the local and global correlations in a cosegmentation system and expand the application of sparse and low-rank represented model, providing a theory basis and technique support for subsequent image and video processing work.

英文关键词： Cosegmentation;Hierarchy Graph ;Structured Sparsity;Low-rank