时空结合相机关联自适应特征增强的视频行人再识别算法研究

针对行人姿势、外部遮挡、光照强度和摄像设备等内外部条件变化导致的行人再识别率较低的问题,提出时空特征结合相机关联自适应特征增强-MFA的视频行人再识别算法。本文首先基于视频提取时空梯度方向直方图(HOG3D)特征,基于图像提取表观特征,然后将两者结合作为视频行人目标的特征描述子,从而提高特征描述有效性;距离度量时将特征进行自适应特征增强后再作边际费希尔分析(Marginal Fisher Analysis, MFA),增强共性特征之间的联系,进一步提高距离度量阶段对特征的判别性。基于iLIDS-VID 和PRID 2011两大视频行人数据集讨论加入时空梯度方向直方图特征和相机关联自适应特征增强的算法性能提升,多组实验结果表明,该算法能够充分利用视频中包含的运动信息,得到鲁棒的视频行人再识别匹配模型,提高行人再识别的匹配精度。      

采用长度规整MAP的说话人分割聚类

本文首次提出了长度规整的最大后验估计(MAP)方法,并将其应用到说话人分割聚类中的交叉似然比(CLR)和T Test这两种度量距离上。传统的MAP方法需要在通用背景模型(UBM)基础上进行统计量的计算,进而对模型参数进行自适应偏移,因此偏移的程度与语音片段的长度正相关。当在度量两个长度不相同的语音片段的相似性时,传统的MAP方法会使得说话人模型刻画不准确,从而影响距离度量。本文在MAP过程中,根据语音的长度对相关因子进行规整,然后再进行模型参数的调整,从而使得模型参数与语音长度无关,更能体现说话人的身份信息。在中文多人电视访谈节目数据的分割聚类评测任务上,采用长度规整的MAP方法相对于传统方法都有明显提升,在CLR度量准则下分割聚类错误率相对下降了35%,在T Test度量准则下分割聚类错误率相对下降了107%。      

平移变化性相似性学习的行人重识别算法

行人重识别的精确度主要取决于相似性度量方法和特征学习模型。现有的度量方法存在平移不变性的特点,会增加网络参数训练的难度。现有的几种特征学习模型只强调样本之间的绝对距离而忽略了正样本对和负样本对之间的相对距离,造成网络学习到的特征判别性不强。针对现有度量方法的缺点该文提出一种平移变化的距离度量方法,能够简化网络的优化并能高效度量图像之间的相似性。针对特征学习模型的不足,提出一种增大间隔的逻辑回归模型,模型通过增大正负样本对之间的相对距离,使得网络得到的特征判别性更强。实验中,在Market1501和CUHK03数据库上对所提度量方式和特征学习模型的有效性进行验证,实验结果表明,所提度量方式性能更好,其平均精确率超出马氏距离度量6.59%,且所提特征学习模型也取得了很好的性能,算法的平均精确率较现有的先进算法有显著提高。     

基于对称KL距离的用户行为时序聚类方法

网络用户随时间变化的行为分析是近年来用户行为分析的热点,通常为了发现用户行为的特征需要对用户做聚类处理。针对用户时序数据的聚类问题,现有研究方法存在计算性能差,距离度量不准确的缺点,无法处理大规模数据。为了解决上述问题,该文提出基于对称KL距离的用户行为时序聚类方法。首先将时序数据转化为概率模型,从划分聚类的角度出发,在距离度量中引入KL距离,用以衡量不同用户间的时间分布差异。针对实网数据中数据规模大的特点,该方法在聚类的各个环节针对KL距离的特点做了优化,并证明了一种高效率的聚类质心求解办法。实验结果证明,该算法相比采用欧式距离和DTW距离度量的聚类算法能提高4%的准确度,与采用medoids聚类质心的聚类算法相比计算时间少了一个量级。采用该算法对实网环境中获取的用户流量数据处理证明了该算法拥有可行的应用价值。     

Towards self-similarity consistency and feature discrimination for unsupervised domain adaptation

Recent advances in unsupervised domain adaptation mainly focus on learning shared representations by global statistics alignment, such as the Maximum Mean Discrepancy (MMD) which matches the Mean statistics across domains. The lack of class information, however, may lead to partial alignment (or even misalignment) and poor generalization performance. For robust domain alignment, we argue that the similarities across different features in the source domain should be consistent with that in the target domain. Based on this assumption, we propose a new domain discrepancy metric, i.e., Self-similarity Consistency (SSC), to enforce the pairwise relationship between different features being consistent across domains. The Gram matrix matching and Correlation Alignment is proven to be a special case, and a sub-optimal measure of our proposed SSC. Furthermore, we also propose to mitigate the side effect of the partial alignment and misalignment by incorporating the discriminative information of the deep representations. Specifically, a simple yet effective feature norm constraint is exploited to enlarge the discrepancy of inter-class samples. It relieves the requirements of strict alignment when performing adaptation, therefore improving the adaptation performance significantly. Extensive experiments on visual domain adaptation tasks demonstrate the effectiveness of our proposed SSC metric and feature discrimination approach.     

Rate adaptation for dynamic adaptive streaming over HTTP in content distribution network

Recently the 3rd Generation Partnership Project (3GPP) and the Moving Picture Experts Group (MPEG) specified Dynamic Adaptive Streaming over HTTP (DASH) to cope with the shortages in progressive HTTP based downloading and Real-time Transport Protocol (RTP) over the User Datagram Protocol (UDP), shortly RTP/UDP, based streaming. This paper investigates rate adaptation for the serial segment fetching method and the parallel segment fetching method in Content Distribution Network (CDN). The serial segment fetching method requests and receives segments sequentially whereas the parallel segment fetching method requests media segments in parallel. First, a novel rate adaptation metric is presented in this paper, which is the ratio of the expected segment fetch time (ESFT) and the measured segment fetch time to detect network congestion and spare network capacity quickly. ESFT represents the optimum segment fetch time determined by the media segment duration multiplied by the number of parallel HTTP threads to deliver media segments and the remaining duration to fetch the next segment to keep a certain amount of media time in the client buffer. Second, two novel rate adaptation algorithms are proposed for the serial and the parallel segment fetching methods, respectively, based on the proposed rate adaptation metric. The proposed rate adaptation algorithms use a step-wise switch-up and a multi-step switch-down strategy upon detecting the spare networks capacity and congestion with the proposed rate adaptation metric. To provide a good convergence in the representation level for DASH in CDN, a sliding window is used to measure the latest multiple rate adaptation metrics to determine switch-up. To decide switch-down, a rate adaptation metric is used. Each rate adaptation metric represents a reception of a segment/portion of a segment, which can be fetched from the different edge servers in CDN, hence it can be used to estimate the corresponding edge server bandwidth. To avoid buffer overflow due to a slight mismatch in the optimum representation level and bandwidth, an idling method is used to idle a given duration before sending the next segment. In order to solve the fairness between different clients who compete for bandwidth, the prioritized optimum segment fetch time is assigned to the newly joined clients. The proposed rate adaptation method does not require any transport layer information, which is not available at the application layer without cross layer communication. Simulation results show that the proposed rate adaptation algorithms for the serial and the parallel segment fetching methods quickly adapt the media bitrate to match the end-to-end network capacity, provide an advanced convergence and fairness between different clients and also effectively control buffer underflow and overflow for DASH in CDN. The reported simulation results demonstrate that the parallel rate adaptation outperforms the serial DASH rate adaptation algorithm with respect to achievable media bitrates while the serial rate adaptation is superior to the parallel DASH with respect to the convergence and buffer underflow frequency.     

A new sequential decoding algorithm based on branch metric

A new sequential decoding algorithm with an adjustable threshold and a new method of moving through the decoding tree is proposed. Instead of the path metric of the conventional sequential decoding algorithms, the proposed algorithm uses a branch metric based on maximum-likelihood criterion. Two new parameters, the jumping-back distance and going-back distance, are also introduced. The performance of the algorithm for long constraint length convolutional codes is compared to those of the other sequential decoding algorithms and the Viterbi algorithm. The results show that the proposed algorithm is a good candidate for decoding of convolutional codes due to its fast decoding capability and good bit error rate (BER) performance. This work was supported in part by the Research Foundation at Karadeniz Technical University under Grant 2004.112.004.01 and 2005.112.009.2.     

深层融合度量子空间学习稀疏特征提取算法

特征提取作为模式识别中的重要步骤,一直是图像处理研究的重点,逐渐兴起的深度学习理论,作为一种新的深层特征提取模型,越来越受到广大学者的关注。本文提出一种基于深层融合度量学习的稀疏特征提取算法,在深度学习的框架内,构建度量映射矩阵,对图像进行分层映射,最大化保留样本集类间区分信息,并且通过稀疏迭代来保证特征提取结果的稀疏性。首先构建图像集距离度量函数,然后通过求解最大化类间距离来确定最优度量映射矩阵,同时对特征映射结果进行 范数稀疏迭代,提高噪声鲁棒性。然后对这个基本特征提取单元进行深度化改造,在第二层中进行同样的度量学习操作,最终通过多层融合提取得到分层深度稀疏特征。相对于已有子空间方法,本文在特征映射过程中引入度量自学习机制,并着重对各个特征映射层进行视觉合理性稀疏约束,融合多层特征语义描述生成最终特征提取结果。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。      

特征压缩在线距离度量学习跟踪

为提高在线学习目标跟踪的实时性和准确率,结合压缩感知理论,提出一种将距离度量学习(DML)运用到目标跟踪的算法。首先,根据所选定的目标位置分别提取目标和背景样本集,运用随机投影理论对样本的Harr-like特征进行压缩;然后,用压缩后的低维特征向量集训练度量矩阵;最后,在新的一帧中抽取目标和背景的样本,用训练得到的度量矩阵计算已知目标和样本间的Mahalanobis距离,距离最小的样本的位置就是所要跟踪的目标的位置。对不同视频序列的测试结果表明,用压缩特征表示目标,使特征计算的计算量压缩到原来的1/4,减少了特征计算的时间;用训练后的度量矩阵计算目标位置,即跟踪器能够根据目标的不断变化自适应调整参数,提高了跟踪的准确率。     

Channel quality estimation and rate adaptation for cellular mobileradio

We propose a technique to measure channel quality in terms of signal-to-interference plus noise ratio (SINR) for the transmission of signals over fading channels. The Euclidean distance (ED) metric, associated with the decoded information sequence or a suitable modification thereof, is used as a channel quality measure. Simulations show that the filtered or averaged metric is a reliable channel quality measure which remains consistent across different coded modulation schemes and at different mobile speeds. The average scaled ED metric can be mapped to the SINR per symbol. We propose the use of this SINR estimate for data rate adaptation, in addition to mobile assisted handoff (MAHO) and power control. We particularly focus on data rate adaptation and propose a set of coded modulation schemes which utilize the SINR estimate to adapt between modulations, thus improving the data throughput. Simulation results show that the proposed metric works well across the entire range of Dopplers to provide near-optimal rate adaptation to average SINR. This method of adaptation averages out short-term variations due to Rayleigh fading and adapts to the long-term effects such as shadowing. At low Dopplers, the metric can track Rayleigh fading and match the rate to a short-term average of the SINR, thus further increasing throughput     

Object detection using multi-resolution mosaic in image sequences

Object detection in image sequences has a very important role in many applications such as surveillance systems, tracking and recognition systems, coding systems and so on. This paper proposes a unified framework for background subtraction, which is very popular algorithm for object detection in image sequences. And we propose an algorithm using spatio-temporal thresholding and truncated variable adaptation rate (TVAR) for object detection and background adaptation, respectively. Especially when the camera moves and zooms in on something to track the target, we generate multi-resolution mosaic which is made up of many background mosaics with different resolution, and use it for object detection. Some experimental results in various environments show that the averaged performance of the proposed algorithm is good.     

基于视觉模糊的LBP鲁棒特征提取与匹配

目的针对目前模糊图像特征提取与匹配方面, 存在特征提取困难、匹配率低、抗噪以及抗尺度变 化能力弱的缺陷。方法提出一种基于SIFT算法与改进的中心对称局部二值模式相结合的精准 、特征识别 率高的匹配算法。首先采用SIFT进行特征的提取,生成多维的描述子,其次采用本文改进的 中心对称局 部二值模式对高维特征描述子进行降维处理,并采用局部特征区域对降维后的描述子进行特 征检测,并生 成纹理特征图像以及信息分布直方图,对特征区域的特征点进行信息量统计,并设置检测阈 值。提取符合 特征信息要求的特征点,并依据Hausdorff距离算法实现图像粗匹配,最后采用RANSAC算法 进行误差匹 配的剔除来改善匹配的精度和鲁棒性。结果测试结果表明,本文所建议的算法是有效的,它 不仅具有良 好的模糊图像分辨能力和抗尺度变化特性,而且具有较强的噪声抑制能力和抗光照变化能力 。结论本文 提出的基于视觉模糊的鲁棒特征匹配算法,不仅考虑到传统特征匹配算法的优缺点,也提出 了算法改进的 新思路,而且较SIFT算法以及LBP算法稳定性和准确度有了明显的提高。     

Data Clustering Method Using a Modified Gaussian Kernel Metric and Kernel PCA

Most hyper‐ellipsoidal clustering (HEC) approaches use the Mahalanobis distance as a distance metric. It has been proven that HEC, under this condition, cannot be realized since the cost function of partitional clustering is a constant. We demonstrate that HEC with a modified Gaussian kernel metric can be interpreted as a problem of finding condensed ellipsoidal clusters (with respect to the volumes and densities of the clusters) and propose a practical HEC algorithm that is able to efficiently handle clusters that are ellipsoidal in shape and that are of different size and density. We then try to refine the HEC algorithm by utilizing ellipsoids defined on the kernel feature space to deal with more complex‐shaped clusters. The proposed methods lead to a significant improvement in the clustering results over K‐means algorithm, fuzzy C‐means algorithm, GMM‐EM algorithm, and HEC algorithm based on minimum‐volume ellipsoids using Mahalanobis distance.     

Spatio-temporal metric learning for individual recognition from locomotion

Individual recognition from locomotion is a challenging task owing to large intra-class and small inter-class variations. In this article, we present a novel metric learning method for individual recognition from skeleton sequences. Firstly, we propose to model articulated body on Riemannian manifold to describe the essence of human motion, which can reflect biometric signatures of the enrolled individuals. Then two spatia-temporal metric learning approaches are proposed, namely Spatio-Temporal Large Margin Nearest Neighbor (ST-LMNN) and Spatio-Temporal Multi-Metric Learning (STMM), to learn discriminant bilinear metrics which can encode the spatio-temporal structure of human motion. Specifically, the ST-LMNN algorithm extends the bilinear model into classical Large Margin Nearest Neighbor method, which learns a low-dimensional local linear embedding in the spatial and temporal domain, respectively. To further capture the unique motion pattern for each individual, the proposed STMM algorithm learns a set of individual-specific spatio-temporal metrics, which make the projected features of the same person closer to its class mean than that of different classes by a large margin. Beyond that, we present a new publicly available dataset for locomotion recognition to evaluate the influence of both internal and external covariant factors. According to the experimental results from the three public datasets, we believe that the proposed approaches are both able to achieve competitive results in individual recognition.     

On design of error-correcting reversible variable length codes

We propose an algorithm for construction of reversible variable length codes (RVLCs) with good error-correcting properties. The error-correcting properties are evaluated by a metric called the free distance, which is always greater than one in the case of the proposed RVLCs. Since variable length codes (VLCs) typically have free distance equal to one, the proposed RVLCs exhibit significant improvement in symbol error rate relative to VLCs constructed using standard methods.     

一种基于距离调节的聚类算法

针对k-means算法不适合凹形样本空间的问题,提出了一种基于距离调节的聚类算法.算法中引入了一种调节最短路径距离作为算法的相似度函数,该函数可以使经过高密度数据区域的两点距离缩短,而经过低密度数据区域的两点距离加长,由此来缩小类间样本的相似度,同时加大类间的相似度,以及更好的聚类.实验结果证明,该算法对凹状的聚类样本空间具有很好的聚类效果.     

原空间中的核SOM分类器

自组织特征映射(SOM)是Kohonen提出的一种人工神经网络模型,其整个学习过程是在输入样本空间内进行,并以欧氏距离为度量.这将导致当输入样本分布结构呈高度非线性时,其分类能力下降.核方法通过核函数实现了一个从低维输入空间到高维特征空间的映射,从而使输入空间中复杂的样本结构在特征空间中变得简单.Donald等人通过核映射将低维输入空间中的非线性问题变换至高维特征空间中,从而使SOM聚类形成于映射后的高维特征空间中.但其缺点是失去了对原输入空间聚类中心及结果的直观刻画;本文采用核方法的目的是为原输入空间诱导出一类异于欧氏距离的新的距离度量,并使原SOM成为特例.而核的多样性进一步可诱导出原空间中不同的度量,导致各种对应SOM分类器的生成.最后,本文侧重通过几种经典的核函数在Benchmark上的试验,对该分类器的性能及可靠性进行了验证.     

SFR-Net:sample-aware and feature refinement network for cross-domain micro-expression recognition

Over the past several decades, micro-expression recognition (MER) has become a growing concern for scientific community. As the filming conditions vary from database to database, previous single-domain MER methods generally exhibit severe performance drop when applied to another database. To deal with this pressing problem, in this paper, a sample-aware and feature refinement network (SFR-Net) is proposed, which combines domain adaptation with deep metric learning to extract intrinsic features of micro-expressions for accurate recognition. With the help of decoders, siamese networks increasingly refine shared features relevant to emotions while exclusive features irrelevant to emotions are gradually obtained by private networks. In order to achieve promising performance, we further design sample-aware loss to constrain the feature distribution in the high-dimensional feature space. Experimental results show the proposed algorithm can effectively mitigate the diversity among different micro-expression databases, and achieve better generalization performance compared with state-of-the-art methods.     

FPI-PD-MCC：一种基于模糊PI-PD的组播拥塞控制算法

针对TFMCC算法速率振荡大的局限性,提出了一种基于模糊PI—PD的组播拥塞控制算法（FPI—PD—MCC：Fuzzy Logic—based Proportional Integral-proportional Derivatire Multicast Congestion Control Algorithm）。在FPI—PD—MCC中,对发送方的速率调整步长进行了平滑,在路由器中引入了PI控制,并利用模糊逻辑计算参数α,从而自动调节丢包概率以缓解拥塞。仿真结果表明,该算法能够使队列长度稳定在期望值附近,同时保证网络吞吐量的平缓变化。