基于Manifold Ranking和结合前景背景特征的显著性检测

针对基于图和流形排序(Manifold Ranking)的显著性检测算法(MR算法)过度依赖边界节点的背景特征的问题,提出一种改进的结合前景背景特征的显著性检测算法。首先,对图像进行超像素分割,建立闭环图模型;然后利用流形排序算法根据图像前景特征和背景特征分别得出前景种子和背景种子;再通过亮度和颜色特征对两类种子进行结合,筛选出更为准确的查询节点;最后再利用流形排序算法进行显著值计算,得到最终的显著图。实验表明,改进方法与MR算法相比在精确率、召回率、F值等多个评价指标上均有明显提升,得到的显著图更接近真值。

Saliency Rank:Two-stage manifold ranking for salient object detection

Salient object detection remains one of the most important and active research topics in computer vision,with wide-ranging applications to object recognition,scene understanding,image retrieval,context aware image editing,image compression,etc. Most existing methods directly determine salient objects by exploring various salient object features.Here,we propose a novel graph based ranking method to detect and segment the most salient object in a scene according to its relationship to image border(background) regions,i.e.,the background feature.Firstly,we use regions/super-pixels as graph nodes,which are fully connected to enable both long range and short range relations to be modeled. The relationship of each region to the image border(background) is evaluated in two stages:(i) ranking with hard background queries,and(ii) ranking with soft foreground queries. We experimentally show how this two-stage ranking based salient object detection method is complementary to traditional methods,and that integrated results outperform both. Our method allows the exploitation of intrinsic image structure to achieve high quality salient object determination using a quadratic optimization framework,with a closed form solution which can be easily computed.Extensive method evaluation and comparison using three challenging saliency datasets demonstrate that our method consistently outperforms 10 state-of-theart models by a big margin.

Syndrome Classification Based on Manifold Ranking for Viral Hepatitis

Treatment determination based on syndrome differentiation is the key of Chinese medicine. A feasible way of improving the clinical therapy effectiveness is needed to correctly differentiate the syndrome classifications based on the clinical manifestations. In this paper, a novel data mining method based on manifold ranking （MR） is proposed to explore the relation between syndromes and symptoms for viral hepatitis. Since MR could take the symptom data with expert differentiation and the symptom data without expert differentiation into the task of syndrome classification, the clinical information used for modeling the syndrome features is greatly enlarged so as to improve the precise of syndrome classification. In addition, the proposed method of syndrome classification could also avoid two disadvantages in previous methods： linear relation of the clinical data and mutually exclusive symptoms among different syndromes. And it could help exploit the latent relation between syndromes and symptoms more effectively. Better performance of syndrome classification is able to be achieved according to the experimental results and the clinical experts.

Audio Retrieval Based on Manifold Ranking and Relevance Feedback

An audio information retrieval model based on Manifold Ranking(MR) is proposed, and ranking results are improved using a Relevance Feedback(RF) algorithm. Timbre components are employed as the model’s main feature. To compute timbre similarity, extracting the spectrum features for each frame is necessary; the large set of frames is clustered using a Gaussian Mixture Model(GMM) and expectation maximization. The typical spectra frame from GMM is drawn as data points, and MR assigns each data point a relative ranking score, which is treated as a distance instead of as traditional similarity metrics based on pair-wise distance. Furthermore, the MR algorithm can be easily generalized by adding positive and negative examples from the RF algorithm and improves the final result. Experimental results show that the proposed approach effectively improves the ranking capabilities of existing distance functions.

Saliency Detection via Manifold Ranking Based on Robust Foreground

The graph-based manifold ranking saliency detection only relies on the boundary background to extract foreground seeds,resulting in a poor saliency detection result,so a method that obtains robust foreground for manifold ranking is proposed in this paper.First,boundary connectivity is used to select the boundary background for manifold ranking to get a preliminary saliency map,and a foreground region is acquired by a binary segmentation of the map.Second,the feature points of the original image and the filtered image are obtained by using color boosting Harris corners to generate two different convex hulls.Calculating the intersection of these two convex hulls,a final convex hull is found.Finally,the foreground region and the final convex hull are combined to extract robust foreground seeds for manifold ranking and getting final saliency map.Experimental results on two public image datasets show that the proposed method gains improved performance compared with some other classic methods in three evaluation indicators:precision-recall curve,F-measure and mean absolute error.

Robust interactive image segmentation via graph-based manifold ranking

Interactive image segmentation aims at classifying the image pixels into foreground and background classes given some foreground and background markers. In this paper, we propose a novel framework for interactive image segmentation that builds upon graph-based manifold ranking model, a graph-based semi-supervised learning technique which can learn very smooth functions with respect to the intrinsic structure revealed by the input data. The final segmentation results are improved by overcoming two core problems of graph construction in traditional models: graph structure and graph edge weights. The user provided scribbles are treated as the must-link and must-not-link constraints. Then we model the graph as an approximatively k-regular sparse graph by integrating these constraints and our extended neighboring spatial relationships into graph structure modeling. The content and labels driven locally adaptive kernel parameter is proposed to tackle the insufficiency of previous models which usually employ a unified kernel parameter. After the graph construction,a novel three-stage strategy is proposed to get the final segmentation results. Due to the sparsity and extended neighboring relationships of our constructed graph and usage of superpixels, our model can provide nearly real-time, user scribble insensitive segmentations which are two core demands in interactive image segmentation. Last but not least, our framework is very easy to be extended to multi-label segmentation,and for some less complicated scenarios, it can even get the segmented object through single line interaction. Experimental results and comparisons with other state-of-the-art methods demonstrate that our framework can efficiently and accurately extract foreground objects from background.

A RANK THEOREM FOR NONLINEAR SEMI-FREDHOLM OPERATORS BETWEEN TWO BANACH MANIFOLDS

In this article thc concept of local conjugation of a C^1 mapping between two Banach manifolds is introduced. Thcn a rank theorem for nonlinear scmi-Fredholm operators between two Banach manifolds and a finite rank theorem are established in global analysis.

基于子空间字典低秩表示的流形投影学习

低秩表示(Low-Rank Representation,LRR)能够将每个数据点表示为若干个基的线性组合,是一种获取样本底层低维结构的方法。然而,大多数LRR方法使用原始数据集作为字典,这不能揭示数据的真实分割。本文提出了基于子空间字典低秩表示的流形投影学习:该方法学习最优子空间作为LRR问题的字典,而不是使用原始数据集;利用基数最少的方案,低秩表示矩阵能很好地恢复原始数据;通过对投影矩阵施加行稀疏约束,该方法不仅可以选择鉴别性特征并忽略冗余特征,而且使子空间学习具有很好的解释性。此外,通过引入流形结构保持约束,使得样本的原始表示和距离信息在投影下保持不变。在多个真实世界数据集上的实验结果表明,该方法优于最近提出的一些相关方法。

采用Ranking Saliency算法改进的交通标志检测方法

目的:交通标志的检测是车载辅助系统的关键环节之一,针对YOLO V3检测算法得到的检测结果存在目标框不精确的问题,提出改进交通标志目标检测算法。方法:YOLO V3是当前目标检测算法中检测召回率高且速度较优的算法,但在定位上不够准确。为解决该问题,本文采用流行排序算法对得到的检测框进行二次修正,从而使得目标定位精度提升。结果:通过结合YOLO V3和流行排序算法使目标检测框的交并比提升了3%~9%。结论:通过YOLO V3和Ranking Saliency的结合能够使得目标检测的定位精度提高。

结合共享近邻和流形距离的自适应谱聚类算法

谱聚类算法是建立在图论的基础上,将聚类问题转化为图的划分问题,能识别任意形状的类簇且易于实现,因此比传统聚类算法具有更强的适应性。然而,该算法中常用的距离度量不能同时考虑全局和局部一致性,且易受到噪声影响;聚类结果依赖由输入数据构造的相似度矩阵,且通过特征分解得到松弛划分矩阵和离散化过程的两步独立策略难以得到一个共同最优解。因此,提出一种结合共享近邻和流形距离的自适应谱聚类算法(SNN-MSC),引入一种新的具有指数项和比例因子的流形距离,可以灵活调整同一流形内数据的相似度和不同流形之间数据的相似度之比,并将密度因子纳入流形距离度量中,以消除噪声影响;采用共享近邻重新定义相似度度量,能挖掘数据点之间的空间结构和局部关系;同时,对拉普拉斯矩阵施加秩约束,使相似度矩阵中的连通分量完全等于簇个数,能够在优化求解过程中自适应优化数据相似度矩阵和聚类结构,无须再进行离散化操作。在人工数据集和UCI真实数据集上的对比实验显示,所提算法在多个聚类有效性指标上能体现出更好的性能。

基于改进局部保持投影算法的轴承故障特征提取

在局部保持投影算法中常采用欧氏距离来度量线性数据间的相似性,而在实际工程中欧氏距离并不能准确度量非线性数据间的相似性,从而使算法对高维数据的特征提取性能不佳。针对以上问题提出了基于相关性Rank-order度量的局部保持投影算法,首先应用相关系数计算样本间相关距离实现数据相似性的初步度量,在此基础上再计算Rank-order距离用于评估具有共享邻居的样本间相似性从而得到更有效的相似度矩阵并构建降维模型,最后通过在公开轴承故障数据集和实际采集的轴承数据集上的实验表明,与传统方法相比改进算法对非线性数据有较好的特征提取效果,证明了改进算法的有效性。

Fourth Rank Energy-Momentum Tensor

In this work, we introduce the new concept of fourth rank energy-momentum tensor. We first show that a fourth rank electromagnetic energy-momentum tensor can be constructed from the second rank electromagnetic energy-momentum tensor. We then generalise to construct a fourth rank stress energy-momentum tensor and apply it to Dirac field of quantum particles. Furthermore, since the established fourth rank energy-momentum tensors have mathematical properties of the Riemann curvature tensor, thus it is reasonable to suggest that quantum fields should also possess geometric structures of a Riemannian manifold.

基于对偶流形重排序的无监督特征选择算法

在许多数据分析任务中,经常会遇到高维数据。特征选择技术旨在从原始高维数据中找到最具代表性的特征,但由于缺乏类标签信息,相比有监督场景,在无监督学习场景中选择合适的特征困难得多。传统的无监督特征选择方法通常依据某些准则对样本的特征进行评分,在这个过程中样本是被无差别看待的。然而这样做并不能完全捕捉数据的内在结构,不同样本的重要性应该是有差异的,并且样本权重与特征权重之间存在一种对偶关系,它们会互相影响。为此,提出了一种基于对偶流形重排序的无监督特征选择算法(Unsupervised Feature Selection Algorithm based on Dual Manifold Re-Ranking, DMRR),分别构建不同的相似性矩阵来刻画样本与样本、特征与特征、样本与特征的流形结构,并结合样本与特征的初始得分进行流形上的重排序。将DMRR与3种原始无监督特征选择算法以及2种无监督特征选择后处理算法进行比较,实验结果表明样本重要性信息、样本与特征之间的对偶关系有助于实现更优的特征选择。

一种稀疏流形低秩表示的子空间聚类方法

针对基于非负低秩稀疏表示的子空间聚类方法不能准确描述数据集结构的问题,提出了一种稀疏流形低秩表示的子空间聚类方法。该方法使用双曲正切函数代替核范数来估计秩函数,并利用加权稀疏正则项使表示系数矩阵稀疏,同时引入稀疏流形正则项来刻画数据集的内在流形结构信息。首先通过带有自适应惩罚的线性交替方向法求解子空间表示模型。然后利用获得的表示系数矩阵构造相似度矩阵,结合使用谱聚类方法得到数据集的聚类结果,最后采用基于局部和全局一致性的半监督分类方法获得数据集的分类结果。在Extended Yale B数据库、CMU PIE数据库、ORL数据库、COIL 20数据库和MNIST数据库上的实验结果表明,本方法可以提高子空间聚类和半监督学习的准确率。

融合局部低秩先验与Bloch流形约束的磁共振指纹重建算法

为了实现快速成像,磁共振指纹(Magnetic Resonance Fingerprinting,MRF)技术通常使用非笛卡尔稀疏采样模板对K空间进行高度欠采样,从而获得稀疏K空间信号.然而,从稀疏的K空间信号重建像空间数据是一个病态不适定问题,重建出的MRF像空间数据存在大量的混叠伪影,直接影响到组织生理参数的重建准确度.为此需要将各种先验知识引入重建模型之中,以缓解MRF重建问题的不适定性.针对上述问题,本文提出一种融合局部低秩先验与Bloch流形约束的MRF重建模型,并使用交替方向乘子法(Alternating Direction Method of Multipliers,ADMM)求解模型中的非凸MRF重建问题.本文算法在引入MRF像空间数据的局部低秩先验的同时,使用预先构建的字典为重建指纹提供流形约束.一方面通过空域局部低秩约束有效抑制混叠伪影的产生,另一方面利用字典先验避免指纹的时域流形特征在迭代重建过程中丢失.仿真实验结果表明,相较于引入了全局低秩先验与Bloch流形约束的其他同类算法,本文算法可以提供更高的组织参数重建准确度.

距离信息与排序模型融合的显著性目标检测方法

针对图像探测系统对地面场景中目标探测及识别问题,提出基于距离轮廓信息与排序模型相融合的显著性目标检测方法,利用二维简易距离轮廓信息对可见光及红外图像实施分割,采用流形排序模型对图像进行处理,进而提高目标的检测性能。实验结果表明,白天条件下对显著性目标检测的准确率可达到87.7%,召回率达到91.6%;夜间条件下的检测准确率达到81.2%,召回率达到86.3%,证明了该方法在保证图像处理速度的前提下,能够显著提高目标检测的正确性。

基于动态规划和流形排序的知识库问答未登录词处理

为解决知识库问答中较多实体未被纳入到词库中以及新词层出不穷旧词库无法及时更新的问题,提出一种基于动态规划和流形排序的知识库问答模型DPQA来为未登录词选择最优表征,缓解知识库未登录词描述信息不足的问题。该方法通过动态规划获取未登录词的子词序列;使用一种基于流形排序的子词排序方法得到最优子词;使用最优子词的向量表示未登录词。实验结果表明,基于动态规划的方法在多个数据集上的问答结果表现最佳。

基于显著性检测的黄瓜叶部病害图像分割算法

针对复杂背景下黄瓜叶部病害分割精度不高的问题,提出了一种基于显著性检测的黄瓜叶部病害图像分割算法。首先利用超像素将黄瓜图像分块,获取黄瓜叶片的边缘,并提出了一种超像素间权重计算方法和显著种子选取方法;然后通过流形排序计算显著图,对得到的显著图进行阈值分割,得到二值图像;再将二值图像与原图像进行掩码运算,得到黄瓜病害叶片;最后利用超绿特征和数学形态学对病害叶片进行分割得到病斑。对常见的黄瓜病害(白粉病、褐斑病、霜霉病、炭疽病)图像进行测试,结果表明该算法与Otsu算法和k-means算法相比,有效解决了冗余分割问题,错分率均在5%以内,算法平均执行时间均小于4 000 ms,分割效果更加精确,为后续构建黄瓜病害自动识别系统奠定了基础。

一种融合图学习与区域显著性分析的图像检索算法

为弥合图像低层视觉特征和高层语义之间的语义鸿沟,改善图像检索的效果,机器学习算法经常被引入到图像检索问题中.通常情况下,机器学习算法是与相关反馈机制相结合,通过用户的交互操作,标定出若干正反例图像,很自然地就可以将图像检索问题转化为模式识别中的分类问题.目前融合区域显著性分析的区域图像检索算法尚没有与机器学习算法相融合.本文结合图像区域显著性分析,并针对用户参与反馈的情况,分别提出了两种图像检索解决方案.其一,在没有用户反馈以及用户只反馈正例图像的情形下,将图像检索问题转化为直推式学习问题(Transductive Learning),改进已有的基于图的半监督学习算法,提出了融合区域显著性分析的层次化图表示(Hierarchi-cal Graph Representation)方式,用以实现标记传播;其二,在用户同时反馈正反例图像的情形下,利用用户反馈得到的正反例图像构建相似性邻接矩阵,通过流形排序算法(Manifold-Ranking)学习出用户感兴趣的查询目标概念并用相应的特征向量集合表示,并据此查询图像库返回用户语义相关的图像集合.实验结果验证了这两种检索策略的有效性.

基于凸包改进的流行排序显著性检测

针对传统的基于图的流行排序显著性检测算法仅仅依赖边界背景先验显著图来提取前景种子,影响最后的排序结果,使得显著性检测结果较差的问题,提出结合凸包提取更精确的前景种子进行流行排序的算法.首先提取图像边界结点作为背景种子进行流行排序得到背景估计显著图,并将该显著图二值化得到粗略的前景区域;然后通过颜色增强的Harris角点检测算法获得图像角点,并用其构造粗略包含显著目标的凸包;最后将凸包和前景区域相结合提取更精确的前景种子进行流行排序得到最后的显著图.在3个公开的图像数据集上,与其他经典算法相比,该算法在PR曲线、MAE值和F-measure上均获得了提升.