兼顾特征级和决策级融合的场景分类

针对单一特征在场景分类中精度不高的问题,借鉴信息融合的思想,提出了一种兼顾特征级融合和决策级融合的分类方法。首先,提取图像的尺度不变特征变换词包(SIFT-BoW)、Gist、局部二值模式(LBP)、Laws纹理以及颜色直方图五种特征。然后,将每种特征单独对场景进行分类得到的结果以Dezert-Smarandache理论(DSmT)推理的方式在决策级进行融合,获得决策级融合下的分类结果;同时,将五种特征串行连接实现特征级融合并进行分类,得到特征级融合下的分类结果。最后,将特征级和决策级的分类结果进行自适应的再次融合完成场景分类。在决策级融合中,为解决DSmT推理过程中基本信度赋值(BBA)构造困难的问题,提出一种利用训练样本构造后验概率矩阵来完成基本信度赋值的方法。在21类遥感数据集上进行分类实验,当训练样本和测试样本各为50幅时,分类精度达到88.61%,较单一特征中的最高精度提升了12.27个百分点,同时也高于单独进行串行连接的特征级融合或DSmT推理的决策级融合的分类精度。     

A hybrid classifier combining SMOTE with PSO to estimate 5-year survivability of breast cancer patients

In this study, we propose a set of new algorithms to enhance the effectiveness of classification for 5-year survivability of breast cancer patients from a massive data set with imbalanced property. The proposed classifier algorithms are a combination of synthetic minority oversampling technique (SMOTE) and particle swarm optimization (PSO), while integrating some well known classifiers, such as logistic regression, C5 decision tree (C5) model, and 1-nearest neighbor search. To justify the effectiveness for this new set of classifiers, the g-mean and accuracy indices are used as performance indexes; moreover, the proposed classifiers are compared with previous literatures. Experimental results show that the hybrid algorithm of SMOTE + PSO + C5 is the best one for 5-year survivability of breast cancer patient classification among all algorithm combinations. We conclude that, implementing SMOTE in appropriate searching algorithms such as PSO and classifiers such as C5 can significantly improve the effectiveness of classification for massive imbalanced data sets.     

An investigation of the effects of correlation and autocorrelation on classifier fusion and optimal classifier ensembles

The purpose of this research was to study various fusion strategies where the levels of correlation between features and auto-correlation within features could be controlled. The fusion strategies were chosen to reflect decision-level fusion (ISOC and ROC), feature level fusion, via a single Generalized Regression Neural Network (GRNN) employing all available features, and an intermediate level of fusion that employed the outputs of individual classifiers, in this case posterior probability estimates, before they are subjected to thresholds and mapped into decisions. This latter scheme involved fusing the posterior probability estimates by employing them as features in a probabilistic neural network. Correlation was injected into the data set both within a feature set (auto-correlation) and across feature sets, and sample size was varied for a two class problem. The fusion methods were then extended to three classifiers, and a method is demonstrated that selects the optimal classifier ensemble.     

语种辨识的多特征信息应用

提取动态的高层语言学特征建立了改进的语种相关的、联合的GMM-LM语种辨识方案。该方案减小了不同语种的高斯混合模型和语言模型之间的相关性,也降低了训练的复杂度。还提出了基于特征提取层和判决层融合技术的语种辨识系统。该系统利用了不同类型的特征对区分不同语种的贡献来增加不同语种语料之间的差异,并使相同语种的语料之间的差异减小。实验表明,设计的语种辨识系统具有较好的扩展性;基于特征提取层和判决层的融合系统能够有效地提高系统识别率。     

Feature-level and decision-level fusion of noncoincidently sampledsensors for land mine detection

We present and compare methods for feature-level (predetection) and decision-level (postdetection) fusion of multisensor data. This study emphasizes fusion techniques that are suitable for noncommensurate data sampled at noncoincident points. Decision-level fusion is most convenient for such data, but it is suboptimal in principle, since targets not detected by all sensors will not obtain the full benefits of fusion. A novel algorithm for feature-level fusion of noncommensurate, noncoincidently sampled data is described, in which a model is fitted to the sensor data and the model parameters are used as features. Formulations for both feature-level and decision-level fusion are described, along with some practical simplifications. A closed-form expression is available for feature-level fusion of normally distributed data and this expression is used with simulated data to study requirements for sample position accuracy in multisensor data. The performance of feature-level and decision-level fusion algorithms are compared for experimental data acquired by a metal detector, a ground-penetrating radar, and an infrared camera at a challenging test site containing surrogate mines. It is found that fusion of binary decisions does not perform significantly better than the best available sensor. The performance of feature-level fusion is significantly better than the individual sensors, as is decision-level fusion when detection confidence information is also available (“soft-decision” fusion)     

Automatic identification of discourse markers in dialogues: An in-depth study of like and well

The lexical items like and well can serve as discourse markers (DMs), but can also play numerous other roles, such as verb or adverb. Identifying the occurrences that function as DMs is an important step for language understanding by computers. In this study, automatic classifiers using lexical, prosodic/positional and sociolinguistic features are trained over transcribed dialogues, manually annotated with DM information. The resulting classifiers improve state-of-the-art performance of DM identification, at about 90% recall and 79% precision for like (84.5% accuracy, κ = 0.69), and 99% recall and 98% precision for well (97.5% accuracy, κ = 0.88). Automatic feature analysis shows that lexical collocations are the most reliable indicators, followed by prosodic/positional features, while sociolinguistic features are marginally useful for the identification of DM like and not useful for well. The differentiated processing of each type of DM improves classification accuracy, suggesting that these types should be treated individually.     

Genetic programming for epileptic pattern recognition in electroencephalographic signals

This paper reports how the genetic programming paradigm, in conjunction with pattern recognition principles, can be used to evolve classifiers capable of recognizing epileptic patterns in human electroencephalographic signals. The procedure for feature extraction from the raw signal is detailed, as well as the genetic programming system that properly selects the features and evolves the classifiers. Based on the data sets used, two different epileptic patterns were detected: 3 Hz spike-and-slow-wave-complex (SASWC) and spike-or-sharp-wave (SOSW). After training, classifiers for both patterns were tested with unseen instances, and achieved sensibility = 1.00 and specificity = 0.93 for SASWC patterns, and sensibility = 0.94 and specificity = 0.89 for SOSW patterns. Results are very promising and suggest that the methodology presented can be applied to other pattern recognition tasks in complex signals.     

Fuzzy integral based information fusion for classification of highly confusable non-speech sounds

Acoustic event classification may help to describe acoustic scenes and contribute to improve the robustness of speech technologies. In this work, fusion of different information sources with the fuzzy integral (FI), and the associated fuzzy measure (FM), are applied to the problem of classifying a small set of highly confusable human non-speech sounds. As FI is a meaningful formalism for combining classifier outputs that can capture interactions among the various sources of information, it shows in our experiments a significantly better performance than that of any single classifier entering the FI fusion module. Actually, that FI decision-level fusion approach shows comparable results to the high-performing SVM feature-level fusion and thus it seems to be a good choice when feature-level fusion is not an option. We have also observed that the importance and the degree of interaction among the various feature types given by the FM can be used for feature selection, and gives a valuable insight into the problem.     

Spectral entropy feature subset selection using SEPCOR to detect alcoholic impact on gamma sub band visual event related potentials of multichannel electroencephalograms (EEG)

The problem of analyzing and identifying regions of high discrimination between alcoholics and controls in a multichannel electroencephalogram (EEG) signal is modeled as a feature subset selection technique that can improve the recognition rate between both groups. Several studies have reported efficient detection of alcoholics by feature extraction and selection in gamma band visual event related potentials (ERP) of a multichannel EEG signal. However, in these studies the correlation between features and their class information is not considered for feature selection. This may lead to redundancy in the feature set and result in over fitting. Therefore in this study, a statistical feature selection technique based on Separability & Correlation analysis (SEPCOR) is proposed to select an optimal feature subset automatically that possesses minimum correlation between selected channels and maximum class separation. The optimal feature selection consists of a ranking method that assigns ranks to channels based on a variability measure (V-measure). From the ranked feature set of highly discriminative features, different subsets are automatically selected by heuristically applying a correlation threshold in steps from 0.02 to 0.1. These subsets are applied as input features to multilayer perceptron (MLP) neural network and k-nearest neighbor (k-NN) classifiers to discriminate alcoholic and control visual ERP. Prior to feature selection, spectral entropy features are computed in gamma sub band (30–55 Hz) interval of a 61-channel multi-trial EEG signal with multiple object recognition tasks. Independent Component Analysis (ICA) is performed on raw EEG data to remove eye blink, motion and muscle artifacts. Results indicate that both classifiers exhibit excellent classification accuracy of 99.6%, for a feature subset of 22 optimal channels with correlation threshold of 0.1. In terms of computation time, k-NN classifier outperforms multilayer perceptron-back propagation (MLP-BP) network with 7.93 s whereas MLP network takes 55 s to perform the recognition task with the same accuracy. Compared to feature section methods used in previous studies on the same EEG alcoholic database, there is a significant improvement in classification accuracy based on the proposed SEPCOR method.     

Binary labelling and decision-level fusion

Two binary labelling techniques for decision-level fusion are considered for reducing correlation in the context of multiple classifier systems. First, we describe a method based on error correcting coding that uses binary code words to decompose a multi-class problem into a set of complementary two-class problems. We look at the conditions necessary for reduction of error and introduce a modified version that is less sensitive to code word selection. Second, we describe a partitioning method for two-class problems that transforms each training pattern into a vertex of the binary hypercube. A constructive algorithm for binary-to-binary mappings identifies a set of inconsistently classified patterns, random subsets of which are used to perturb base classifier training sets. Experimental results on artificial and real data, using a combination of simple neural network classifiers, demonstrate improvement in performance for these techniques, the first suitable for k-class problems, k>2 and the second for k=2.     

基于多特征融合的医学图像识别研究*

以医学图像为研究对象,针对任何一类特征都不能很好地表达医学图像的缺点以及进一步提高医学图像的识别率,提出了一种基于特征级数据融合与决策级数据融合相结合的分类方法。实验结果表明,采用特征级数据融合,融合后的特征可以较好地表达医学图像,且减少了后期分类的计算量;采用决策级数据融合,取得了比单个分类器更高的识别率。     

Feature based quality assessment of DNA sequencing chromatograms

Although next generation sequencing applications are getting dominant in molecular genetics, there are still many institutions that want to utilize their legacy sequencers as much as possible. An important concern in sequencing services is the quality of trace files presented to the customers. In this respect, the quality of the trace files should be screened and low quality files should be handled differently before reaching to customers. The quality scores already present in the trace files provide some useful information, however by incorporating auxiliary information we can improve to reliability of these scores. To this end, we used a feature based supervised classification strategy which requires a set of training and testing trace files qualities of which are determined manually. We tested several machine learning algorithms, namely k-nearest neighbors, Naive Bayes, Support Vector Machines and Random Forest, on a public DNA trace repository. Our results indicate that RF method with only 4 simple features provides a classification accuracy rate of 94.68% with a high level of reliability of concurrence (Kappa = 0.8679).     

Gender recognition from visible and thermal infrared facial images

Most present research of gender recognition focuses on visible facial images, which are sensitive to illumination changes. In this paper, we proposed hybrid methods for gender recognition by fusing visible and thermal infrared images. First, the active appearance model is used to extract features from visible images, as well as local binary pattern features and several statistical temperature features are extracted from thermal infrared images. Then, feature selection is performed by using the F-test statistic. Third, we propose using Bayesian Networks to perform explicit and implicit fusion of visible and thermal infrared image features. For explicit fusion, we propose two Bayesian Networks to perform decision-level and feature-level fusion. For implicit fusion, we propose using features from one modality as privileged information to improve gender recognition by another modality. Finally, we evaluate the proposed methods on the Natural Visible and Infrared facial Expression spontaneous database and the Equinox face database. Experimental results show that both feature-level and decision-level fusion improve the gender recognition performance, compared to that achieved from one modality. The proposed implicit fusion methods successfully capture the role of privileged information of one modality, thus enhance the gender recognition from another modality.     

Hierarchical distance learning by stacking nearest neighbor classifiers

We propose a two-layer decision fusion technique, called Fuzzy Stacked Generalization (FSG) which establishes a hierarchical distance learning architecture. At the base-layer of an FSG, fuzzy k-NN classifiers receive different feature sets each of which is extracted from the same dataset to gain multiple views of the dataset. At the meta-layer, first, a fusion space is constructed by aggregating decision spaces of all the base-layer classifiers. Then, a fuzzy k-NN classifier is trained in the fusion space by minimizing the difference between the large sample and N-sample classification error. In order to measure the degree of collaboration among the base-layer classifiers and the diversity of the feature spaces, a new measure called, shareability, is introduced. Shearability is defined as the number of samples that are correctly classified by at least one of the base-layer classifiers in FSG. In the experiments, we observe that FSG performs better than the popular distance learning and ensemble learning algorithms when the shareability measure is large enough such that most of the samples are correctly classified by at least one of the base-layer classifiers. The relationship between the proposed and state-of-the-art diversity measures is experimentally analyzed. The tests performed on a variety of artificial and real-world benchmark datasets show that the classification performance of FSG increases compared to that of state-of-the art ensemble learning and distance learning methods as the number of classes increases.     

Boosting a fast neural network for supervised land cover classification

It is demonstrated that the use of an ensemble of neural networks for routine land cover classification of multispectral satellite data can lead to a significant improvement in classification accuracy. Specifically, the AdaBoost.M1 algorithm is applied to a sequence of three-layer, feed-forward neural networks. In order to overcome the drawback of long training time for each network in the ensemble, the networks are trained with an efficient Kalman filter algorithm. On the basis of statistical hypothesis tests, classification performance on multispectral imagery is compared with that of maximum likelihood and support vector machine classifiers. Good generalization accuracies are obtained with computation times of the order of 1 h or less. The algorithms involved are described in detail and a software implementation in the ENVI/IDL image analysis environment is provided.