多层融合深度局部PCA子空间稀疏优化特征提取模型

子空间方法是主要利用全局信息的经典模式识别方法,随着深度学习思想的引入,局部自学习结构特征模型得到大家的关注.利用深度学习原理,本文提出一种多层融合的深度局部子空间稀疏优化特征自学习抽取模型解决目标识别问题.首先,对训练样本集通过最小化重构误差得到第一层的主成分（Principal Component Analysis,PCA）特征映射矩阵;然后,通过L₁范数约束对特征映射结果进行稀疏优化,提高算法鲁棒性.接着,在第二层映射层以第一层的特征输出为输入,进行同样的特征矩阵学习操作,最终将图像映射至深层PCA子空间;然后,对各个映射层的特征提取结果进行加权融合,进行二值化哈希编码和直方图分块编码,提取图像的深度子空间稀疏特征.在FERET、AR、Yale等经典人脸数据库以及MNIST、CIFAR-10等目标数据库上的实验结果表明,该算法可以取得较高的识别率以及较好的光照、表情、人脸朝向鲁棒性,并且相对于卷积神经网络等深度学习框架具有结构简洁、收敛速度快等优点.     

Unbalanced JPEG image steganalysis via multiview data match

Image steganalysis must address the matter of learning from unbalanced training sets where the cover objects (normal images) always greatly outnumber the stego ones. But the research in unbalanced image steganalysis is seldom seen. This work just focuses on the problem of unbalance JPEG images steganalysis. In this paper, we propose a frame of feature dimension reduction based semi-supervised learning for high-dimensional unbalanced JPEG image steganalysis. Our method uses standard steganalysis features, and selects the confident stego images from the unlabeled examples by multiview match resampling method to rebalance the unbalanced training images. Furthermore, weighted Fisher linear discriminant (WFLD) is proposed to find the proper feature subspace where K-means provides the weight factor for WFLD in return. Finally, WFLD and K-means both work in an iterative fashion until convergence. Experimental results on the MBs and nsF5 steganographic methods show the usefulness of the developed scheme over current popular feature spaces.     

基于链路选择学习算法的无设备目标定位

针对压缩感知框架下无设备目标定位(device-free localization,DFL)的字典失配问题,提出一种基于链路选择学习(link selection learning,LSL)算法的DFL方式.由于传统基于阴影模型的字典无法准确表达接收信号强度(received signal strength,RSS)变化与目标位置间的对应关系,本文算法首先在训练阶段通过字典学习的方式更新初始字典; 同时该算法在更新字典的过程中,仅选取置信区域中的链路参与计算,这样既加速了字典学习过程,提高了算法实时性,又滤除了野值链路的影响.室内外实验结果表明,该方法可以有效地消除现有基于阴影模型字典所带来的模型误差,提高定位精度,同时具有运算速度快的优点.     

Green learning: Introduction,examples and outlook

Rapid advances in artificial intelligence (AI) in the last decade have been largely built upon the wide applications of deep learning (DL). However, the high carbon footprint yielded by larger and larger DL networks has become a concern for sustainability. Furthermore, DL decision mechanism is somewhat obscure in that it can only be verified by test data. Green learning (GL) is being proposed as an alternative paradigm to address these concerns. GL is characterized by low carbon footprints, lightweight model, low computational complexity, and logical transparency. It offers energy-efficient solutions in cloud centers as well as mobile/edge devices. GL also provides a more transparent, logical decision-making process which is essential to gaining people’s trust. Several statistical tools such as unsupervised representation learning, supervised feature learning, and supervised decision learning, have been developed to achieve this goal in recent years. We have seen a few successful GL examples with performance comparable with state-of-the-art DL solutions. This paper introduces the key characteristics of GL, its demonstrated applications, and future outlook.     

多指标混合逐次投影高光谱图像降维算法

降维对于高光谱图像解译具有重要意义。基于二阶统计量分析的经典主成分分析方法在降维过程中会丢失小目标信号。为解决这一问题,本文中引入高阶统计量作为投影指标对主成分分析方法进行拓展,提出了一种基于不同统计量描述的混合逐次投影的高光谱图像降维算法。该方法在保持主成分分析优点的同时,有效结合了非正交向量投影的特点,可以在降维后的低维空间中保留异常信号成分。真实高光谱图像数据的实验结果证明,该方法相对于主成分分析可以提取更加完整的低维信号子空间。     

采用新型纹理特征的2DLDA人脸识别算法

针对现有基于纹理特征的人脸识别算法中纹理特征维数偏大且对噪声较敏感等不足,提出了用于描述人脸图像大尺度局部特征的中心四点二元模式（Center Quad Binary Pattern, C-QBP）和用于描述图像小尺度局部特征的简化四点二元模式（Simplified Quad Binary Pattern, S-QBP）两种互补的新型纹理特征。在此基础上,实现基于新型纹理特征的2DLDA人脸识别算法。首先对人脸图像进行多级分割,再对所产生的图像块提取C-QBP和S-QBP纹理特征,构建纹理特征矩阵。最后,采用2DLDA子空间学习算法实现基于新型纹理特征的人脸识别。实验结果表明,本文所提出的人脸识别算法的识别率明显高于其他基于纹理特征和子空间学习的人脸识别算法。当每一类训练样本数统一设置为5,特征维数为48×4时,在ORL人脸库上,本文所提出的人脸识别算法的识别率达98.68%;在YALE人脸库上,特征维数为48×36时,识别率达99.42%;在FERET人脸库上,特征维数为48×26时,识别率为91.73%。      

Incremental POP Learning

In recently proposed partial oblique projection (POP) learning, a function space is decomposed into two complementary subspaces, so that functions belonging to one of which can be optimally estimated. This paper shows that when the decomposition is specially performed so that the above subspace becomes the largest, a special learning called SPOP learning is obtained and correspondingly an incremental learning is implemented, result of which equals exactly to that of batch learning including novel data. The effectiveness of the method is illustrated by experimental results.     

Incremental POP Learning

In neural network learning, it is practically necessary to retrain a trained network when novel data become available. Retraining from the scratch is highly inefficient in memory and computation while adjusting adaptively the trained result by the new data is a natural avenue and it is called incremental learning[1]. Recently, incremental learning becomes an interest issue in neural networks, and also many methods have been devised[1~3]. These methods are divided into two categories according …     

基于稀疏表示结合流形距离的超球覆盖可拒绝模式识别算法研究

本文构造了一种基于稀疏表示结合流形距离超球覆盖的可拒绝模式识别模型.由于同类样本可以认为分布在同一个非线性流形上,所以在训练学习过程中首先对各类样本空间构建局部线性流形子空间超球覆盖模型,并选择训练样本.这样对于输入的测试模式,即可根据各类的子空间包含边界做出拒识或者接受处理的判决.然后,针对接受的模式,再利用稀疏表示分类器在训练样本空间范围内进行分类判决.在UCI数据库、MNIST手写体数据库、MIT-CBCL人脸识别数据库和CMU AMP人脸表情数据库上的实验结果表明本文的思路合理可行,在实际应用领域具有一定应用价值.     

快速子空间分解方法及其维数的快速估计

提出一种快速子空间分解方法.该方法只需要知道某一期望信号的训练信号就可以实现信号子空间和噪声子空间的快速估计.给出快速确定子空间维数的方法.子空间维数的估计和子空间的快速分解是同时进行的.本文方法只需要多级维纳滤波器的前向递推,不需要估计协方差矩阵和对其作特征值分解,所以运算量可以明显地降低,而估计的性能接近于常规方法.最后,仿真结果验证了方本文法的有效性.     

The fast recursive row-Householder subspace tracking algorithm

We introduce a new sequential algorithm for tracking the principal subspace and, optionally, the r dominant eigenvalues and associated eigenvectors of an exponentially updated covariance matrix of dimension N×N, where N>r. The method is based on an updated orthonormal-square (QS) decomposition using the row-Householder reduction. This new subspace tracker reaches a dominant complexity of only 3Nr multiplications per time update for tracking the principal subspace, which is the lower bound in dominant complexity for an algorithm of this kind. The new method is completely reflection based. An updating of inverse matrices is not used.     

近邻类加权结构稀疏表示图像识别算法

针对如何将近邻、子空间学习与稀疏表示结合起来解决稀疏分类计算量较大的问题。由于子空间中样本的类内散度小,类间散度大,且同类中所有样本对重构的影响相似,因此按类而非样本处理的思想更符合基于类重构误差进行分类的算法要求,为此提出一种基于近邻类加权结构稀疏表示算法用于图像识别。该算法首先利用线性类重构误差选取 个最近邻类,并将其对应的系数作为权值对投影后的近邻类加权,其次在投影子空间上,用 个类的加权训练样本集对测试样本进行结构稀疏表示,最后根据最小类重构误差得出分类结果。在AR,Yale B,MNIST,PIE数据库上的实验结果表明该方法在训练样本数较少的情况下获得较高的识别率且具有一定的鲁棒性。      

Computationally efficient DOA estimation for coprime linear array: a successive signal subspace fitting algorithm

ABSTRACT

In this paper, direction of arrival (DOA) estimation of multiple signals with coprime array is investigated and signal subspace fitting (SSF) method is linked to the coprime array, which achieves a better DOA estimation performance than the traditional uniform array. While the SSF method requires expensive computational cost in the case of multiple signals due to the multidimensional global angular searching, we propose a successive SSF (S-SSF) algorithm from a computationally efficient perspective. In the proposed algorithm, we employ rotational invariance and coprime property to obtain the initial estimates. Then, via a successive scheme, we transform the traditional multidimensional global angular searching problem into one-dimensional partial angular searching one. Consequently, the computational complexity has been significantly reduced. Specifically, the proposed S-SSF algorithm can obtain almost the same DOA estimation performance as SSF but with remarkably lower complexity. Finally, Cramer-Rao Bound (CRB) is provided and numerical simulations demonstrate the effectiveness of the proposed algorithm.     

基于判别性降维的字典学习在人脸识别中的应用

针对目前大多数人脸识别方法只能单独实施降维或者字典学习而不能完全利用训练样本判别信息的问题,提出了基于判别性降维的字典学习方法,通过联合降维与字典学习使得投影矩阵和字典更好地相互拟合,从而可以获得更高效的人脸分类系统。所提方法的有效性在AR及MPIE两大通用人脸数据库上得到了验证,实验结果表明,相比于几种先进的线性表示方法,所提算法取得了更高的识别率,特别当训练样本数很少的时候,识别效果更佳。     

基于PCA网络的快速子空间分解及其维数估计

提出了一种基于主分量分析（PCA）神经网络实现子快速子空间分解及其维数估计的新方法。该方法不需要估计数据协方差矩阵和特征值分解,只需将阵列数据输入到PCA网络,通过网络权值的无监督自组织迭代即可同时完成子空间分解及其维数估计。因此该方法具有运算量小和复杂度低的特点,易于实时处理。计算机仿真验证了该方法的有效性。     

Subspace tracking with adaptive threshold rank estimation

In frequency and direction of arrival (DOA) tracking problems, singular value decomposition (SVD) can be used to track the signal subspace. Typically, for a problem sizen, only a few, sayr dominant eigencomponents need to be tracked, wherern. In this paper we show how to modify the Jacobi-type SVD to track only ther-dimensional signal subspace by forcing the (n-r)-dimensional noise subspace to be spherical. Therby, the computational complexity is brought down fromO(n²) toO(nr) per update. In addition to tracking the subspace itself, we demonstrate how to exploit the structure of the Jacobi-type SVD to estimate the signal subspace dimension via a simple adptive threshold comparison technique. Most available computationally efficient subspace tracking algorithms rely on off-line estimation of the signal subspace dimension, which acts as a bottleneck in real-time parallel implementations. The noise averaged Jacobi-type SVD updating algorithm presented in this paper is capable of simultaneously tracking the signal subspace and its dimension, while preserving both the low computational cost ofO(nr) and the parallel structure of the method, as demonstrated in a systolic implementation. Furthermore, the algorithm tracks all signal singular values. Their squares are estimates of the powers in the orthogonal modes of the signal. Thus, applications of the algorithm are not limited to only DOA and frequency tracking where information about the powers of signal components is not exploited.     

Transfer subspace learning based on structure preservation for JPEG image mismatched steganalysis

In real-world steganalysis applications, the traditional steganalysis methods built by a set of training data coming from a source may be applied to detect data from another new different source. In this case, the steganalyzers will face a serious problem that training data and test data are no longer subjected to the same distribution, and thus the detection performance would degrade rapidly. To address this problem, a novel transfer subspace learning method with structure preservation for image steganalysis is proposed in this paper. It aims to alleviate the mismatch between the training and test data so as to improve the detection performance. Specifically, a discriminant projection matrix is learned for the training and test data such that the projected data of both sets lie in a common subspace where each sample can be linearly reconstructed by a combination of the training data. In this way, the difference between the training and test sets is decreased. Further, in order to preserve the structure information of features in the projection subspace, a Frobenius-norm based regularization term is introduced into the objective function. Moreover, to mitigate the negative impacts of noises and outliers, a structurally sparse error matrix is introduced to model the noise and outlier information. The formulation of the proposed method can be efficiently solved by an alternating optimization algorithm. The extensive experiments compared with prior arts show the validity of the proposed method for JPEG image mismatched steganalysis.     

Video semantic analysis based on structure-sensitive anisotropic manifold ranking

As a major family of semi-supervised learning (SSL), graph-based SSL has recently attracted considerable interest in the machine learning community along with application areas such as video semantic analysis. In this paper, we analyze the connections between graph-based SSL and partial differential equation- (PDE) based diffusion. From the viewpoint of PDE-based diffusion, the label propagation in normal graph-based SSL is isotropic accompanied with distance. However, according to the structural assumption, which is one of the two basic assumptions in graph-based SSL, we need to enhance the label propagation between the samples in the same structure while weakening the counterpart between the samples in different structures. Accordingly, we deduce a novel graph-based SSL framework, named structure-sensitive anisotropic manifold ranking (SSAniMR), from PDE-based anisotropic diffusion. Instead of using Euclidean distance only, SSAniMR takes local structural difference into account to make the label propagation anisotropic, which is intrinsically different from the isotropic label propagation process in general graph-based SSL methods. Experiments conducted on the TREC Video Retrieval Evaluation (TRECVID) dataset show that this approach significantly outperforms existing graph-based SSL methods and is effective for video semantic annotation.     

网络应用流类别不平衡环境下的SSL加密应用流识别关键技术

通过深入研究网络类别不平衡的原因,选择SMOTE（synthetic minority over-sampling technique）过抽样方法对数据集进行预处理,并充分利用特征匹配高准确性的优点识别和分拣出SSL 加密流,进而利用基于互信息最大化的聚类方法和SVM分类方法进一步识别SSL加密应用,这种混合方法有效地结合了静态特征匹配和机器学习方法的优点,达到识别分类方法在准确性和识别速度的均衡。     

基于类别保留投影的基因表达数据特征提取新方法

 从两两样本的类别关系出发,提出一种新的线性鉴别特征提取方法,叫做类别保留投影.相比经典的fisher线性鉴别分析方法,类别保留投影具有最优子空间维数不受样本类别数限制、计算复杂度低的优点.通过对真实基因表达数据进行样本分类识别,证实了本文方法的有效性.并将类别保留投影方法推广到非线性空间,提出核类别保留投影,用于解决非线性特征提取问题,对基因表达数据的实验验证了方法的可行性.