Search

A new speech feature extraction technique called moving average multi directional local features (MA-MDLF) is presented in this paper. This method is based on linear regression (LR) and moving average (MA) in the time–frequency plane. Three-point LR is taken along time axis and frequency axis, and 3 points MA is taken along 45° and 135° in the time–frequency plane. The LR captures the voice onset\offset, formant contour, while the moving average captures the dynamics on time–frequency axes which can be seen as voiceprints. The MA-MDLF performance is compared to commonly used speech features in speaker recognition. The comparison is performed in a speaker recognition system (SRS) for three different conditions, namely clean speech, mobile speech, and cross channel. MA-MDLF has shown better performance than the baseline MFCC, RASTA-PLP and LPCC. In clean and mobile speech, MA-MDLF feature performs the best and also in the cross channel task MA-MDLF performed excellent. We also evaluated the MA-MDLF using three speech databases, namely KSU, LDC Babylon and TIMITdatabases, and found that MA-MDLF outperformed the other commonly used features with speech from all the three databases. The first and second databases are for Arabic speech while third is for English speech.

Unstructured online reviews are undergoing a rather rapid expansion with the development of E-commerce, and they contain sentiment information in which consumers and businesses are very interested. Therefore, effective sentiment classification has become one of the important research topics. Many studies have shown that ensemble learning methods may have great hopeful applicability in sentiment classification tasks. In this paper, we propose a new ensemble learning framework for sentiment classification of Chinese online reviews. First of all, according to the complicated characteristics of Chinese online reviews, we extract Part of Speech Combination Pattern, Frequent Word Sequence Pattern and Order Preserved Submatrix Pattern as the input features. Furthermore, we use the algorithm of Random Subspace based on Information Gain by considering the problem of massive features in the reviews, which can improve the base classifiers simultaneously. Finally, we adopt the algorithm of Constructing Base Classifiers based on Product Attributes to combine the sentiment information of each attribute in a review so as to obtain better performance on sentiment classification. The experimental results show that the proposed ensemble learning framework has significant improvement in sentiment classification of Chinese online reviews.

Electroencephalogram (EEG) is the most important monitoring methodology for the detection of epileptic seizure diseases. In this paper, EEG based epileptic seizure detection is assessed by employing Bern-Barcelona EEG and Bonn University EEG database. The proposed technique contains three major steps: decomposition, feature extraction and classification. Initially, decomposition using variational mode decomposition delivers an effective frequency localization. After decomposition, semantic feature extraction is carried-out by employing differential entropy and peak-magnitude of root mean square ratio for achieving optimal feature subsets and also for the rejection of irrelevant and redundant features. After finding the feature information, a superior classifier named as random forest is employed for classifying the normality and abnormality of seizure. The experimental result shows that the proposed approach distinguishes the normality and abnormality of seizure EEG signals in terms of sensitivity, specificity, accuracy, positive predictive value and negative predictive value with a superior recognition accuracy.

With the fast growing demands in authentication biometric-based authentication system has been widely utilized in many applications which require secured personal identification/verification. In the existing computer systems, there is a possibility of an imposter gaining access when a user session is active and the user moves away from the system. To solve this problem, proposes a continuous user authentication scheme. Continuous authentication (CA) system verifies the user continuously once a person is logged in. CA system prevents the intruders from invoking the system. It passively verifies the system without interrupting the users work progress. In this paper, score level fusion is carried out using optimization that is genetic particle swarm optimization and classifiers lazy classifier—Naive Bayes held for recognition process. The main objective of the proposed method is to fuse the user biometric traits and to accomplish the optimal result in the continuous user authentication. The proposed technique consists of four modules, namely processing module, feature extraction module, fusion module and recognition module. Finally, the proposed fusion method is applied to remote biometric authentication. The implementation is carried out using MATLAB and the evaluation metrics employed are false acceptance rate, false rejection rate and accuracy, sensitivity and specificity. From the results, we can observe that the proposed technique has achieved better performance metrics values for continuous authentication process.

As we know that the nearest neighbor search is a good and effective method for large-scaled image search. This paper mainly focuses on the design and implementation of parallel image retrieval system based on Hadoop architecture and hashing binary code. Firstly, the whole architecture design of the system and the whole flow of image Hash retrieval are introduced. Then, the image feature extraction based on MAPREDUCE parallel data processing framework, as well as the next feature quantization and the construction method of the Hash index table are introduced in details. Finally, through the comparison of retrieval performances on the INRIA data set, it has been proved that the proposed system has better performance than the single node system on both the retrieval speed and the capability of dealing with massive data.

In order to fully utilize potential feature information of RGB-D images, current popular algorithms mainly use convolutional neural network (CNN) to execute both feature extraction and classification. Such methods could achieve impressive results but usually on the basis of an extremely huge and complex network. What’s more, since the fully connected layers in CNN form a classical neural network classifier, which is trained by gradient descent-based implementations, the generalization ability is limited and sub-optimal. To address these problems, this paper introduce a multi-view CNN-SPMP-RNN-ELM (MCSPMPR-ELM) model for RGB-D object recognition, which combines the power of MCSPMPR and fast training of ELM. It uses the MCSPMPR algorithm to extract discriminative features from raw RGB images and depth images separately. Then the abstracted features are fed to a nonlinear ELM classifier, which leads to better generalization performance with faster learning speed. At last, co-training is employed to learn from the unlabeled data using the two distinct feature sets by semi-supervised learning method. Experimental results on widely used RGB-D object datasets show that our method achieves competitive performance compared with other state-of-the-art algorithms specifically designed for RGB-D data.

Distinguishing subsequence patterns mining aims to discover the differences between different categories of sequence databases and to express characteristics of classes. It plays an important role in biomedicine, feature information selection, time-series classification, and other areas. The existing distinguishing subsequence patterns mining only focuses on whether a pattern appears in a sequence, regardless of the number of occurrences of the pattern in the sequence and the proportion of the pattern in the entire sequence database, which affects the discovery of the distinguishing patterns when there are a large number of irrelevant occurrences. Therefore, the nonoverlapping conditional distinguishing subsequence patterns mining algorithm is proposed. In this paper, we focus on the number of nonoverlapping occurrences that effectively reduce the number of irrelevant or redundant occurrences, and in this way, the number of occurrences can be better grasped. At the same time, we use a specially designed data structure, namely, a Nettree, to avoid backtracking. In addition, we use the distinguishing patterns as classification features, and carry out classification experiments on DNA sequences and time-series data with two classes. Extensive experimental results and comparisons demonstrate the efficiency of the proposed algorithm and the correctness of the feature extraction.

The major challenge of feature based blind steganalysers lies in designing effective image features which give true evidence of the stego noise rather than the natural noise present in the images. Hence they report low detection accuracy in real time implementation in spite of employing 100s of features in the process. In this paper, we coin a new paradigm for detecting steganography by examining the task as a three-steps process with the following repercussions: (a) employing curvelet transform denoising as a pre-processing step that produces better stego noise residuals suppressing the natural noise residual rather than a general denoising step before feature extraction, (b) extracting various steganalytic features, both in spatial domain as well transform domain and (c) implementing the system based on an efficient classifier, multi-surface proximal support vector machine ensemble oblique random rotation forest, that provides detection rate superior to other existing classifiers. Extensive experimentation with huge database of clean and steganogram images produced from seven steganographic schemes with varying embedding rates, and using five steganalysers, shows that the proposed paradigm improves the detection accuracy substantially and proves to be a high performance strategy even at low embedding rates. This model can be employed as a preprocessing component for any image steganalyser and high performance accuracy can be obtained.

According to the problem that the elevation data does not reflect the slope and surface roughness of target topography, the preprocessing of topographic elevation data and extraction algorithm of topographic feature are proposed, and the corresponding extraction of topographic feature is done. A terrain risk assessment method is presented based on terrain roughness and slope information fusion, aiming at the problem that terrain roughness and gradient cannot be directly reflected from the terrain elevation data, in this paper. The innovation is that it is the first time that the bilinear interpolation algorithm is applied in preprocess of elevation data and extraction of topographic feature, as well as the terrain roughness and gradient information fusion algorithm are applied to terrain feature extraction and risk assessment for the first time. By simulation and checking calculation of a certain digital topography example, it is proved that the extraction method of topographic information based on elevation data is feasible and reliable. It will provide a new research approach for target information recognition and topography risk assessment accurately.

Due to the complexity issue of the hand gesture recognition feature extraction, for example the variation of the light and background. In this paper, the convolution neural network is applied to the recognition of gestures, and the characteristics of convolution neural network are used to avoid the feature extraction process, reduce the number of parameters needs to be trained, and finally achieve the purpose of unsupervised learning. Error back propagation algorithm, is loaded into the convolution neural network algorithm, modify the threshold and weights of neural network to reduce the error of the model. In the classifier, the support vector machine that is added to optimize the classification function of the convolution neural network to improve the validity and robustness of the whole model.

Most of the advancements are now carried out by interconnecting physical devices with computers; this is what known as Internet of Things (IoT). The major problems facing by blind people fall in the category of navigating through indoor and outdoor environments consisting of various obstacles and recognition of person in front of them. Identification of objects or person only with perceptive and audio information is difficult. An intelligent, portable, less expensive, self-contained navigation and face recognition system is highly demanded for blind people. This helps blind people to navigate with the help of a Smartphone, global positioning system (GPS) and a system equipped with ultrasonic sensors. Face recognition can be done using neural learning techniques with feature extraction and training modules. The images of friends, relatives are stored in the database of user Smartphone. Whenever a person comes in front of the blind user, the application with the help of neural network gives the voice aid to the user. Thus this system can replace the regular imprecise use of guide dogs as well as white sticks to help the navigation and face recognition process for people with impaired vision.In this paper, we have proposed a novel image recognition and navigation system which provides precise and quick messages in the form of audio to visually challenged people so that they can navigate easily. The performance of the proposed method is comparatively analyzed with the help of ROC analysis.

Big data is centered upon the technique of expanding volume of high velocity, intricate along with different kind of data. Organizations that hold vast sum of data deals with new creation of systematic tools intended for large data. The conventional data-intensive business application starts to go down behind the times, on account of the deficient abilities to manage large data volumes, unstructured information, low rate of information retrieval along with complex algorithms. Big data relies upon the data complexity, relatively than the data size only. For resolving this kind of trouble, this paper establishes a mutual refinement technique for big data retrieval to augment the performance. The intended system comprises the system of training and retrieval which is performed consecutively. In training process, initially input data is preprocessed by splitting the data. Then frequency and entropy features are extracted from the preprocessed data. After the feature extraction data is exhibited to the mutual refinement process. In mutual refinement step hash tag graph is generated to train the data and this removes the uncertainty from the data. In retrieval process, the input query data is used for the similarity assessment. Features like frequency and entropy are extracted from the query data. Then the feature value is compared with the hash tag graph. If the feature value is matched then the data is retrieved as of the hash tag graph and the retrieved data is visualized. The proposed technique’s performance is assessed by relating our intended work with the other conventional works. The experimental output exhibits that our intended mutual refinement process augments the system performance process by confiscating the uncertainty comprised in the system. This work offered a unique mutual refinement approach which yields better outcomes for retrieving the big data in a proficient manner. The proposed process retrieving process in big data gives the better performance but in future, experiments can be done on large datasets and some real-time applications to calculate the effectiveness of the proposed method.

Knowledge bases (such as Wikipedia) are valuable resources of human knowledge which have contributed to various of applications. However, their manual maintenance makes a big lag between their contents and the up-to-date information of entities. Cumulative citation recommendation (CCR) concentrates on identifying worthy-citation documents from a large volume of stream data for a given target entity in knowledge bases. Most previous approaches first carefully extract human-designed features from entities and documents, and then leverage machine learning methods such as SVM and Random Forests to filter worthy-citation documents for target entities. There are some problems in handcraft features for entities and documents: (1) It is an empirical process that requires expert knowledge, thus cannot be easily generalized; (2) The effectiveness of humanly designed features has great effect on the performance; (3) The implementation of the feature extraction process is resource dependent and time-consuming. In this paper, we present a Joint Deep Neural Network Model of Entities and Documents for CCR, termed as DeepJoED, to identify highly related documents for given entities with several layers of neurons, by automatically learn feature extraction of the entities and documents, and train the networks in an end-to-end fashion.An extensive set of experiments have been conducted on the benchmark dataset provided in the Text REtrieval Conference (TREC) Knowledge base acceleration (KBA) task in 2012. The results show the model can bring a significant improvement relative to the state-of-the-art results on this dataset in CCR.

Data science augments manual data understanding with machine learning for potential performance increase. In this paper, data science methodology is examined to enhance machine learning application in smartphone based automatic human activity recognition (HAR). Eventually, a modified feature engineering and a novel post-learning data engineering are proposed in the machine learning framework as the alternate of data understanding for an effective HAR. The proposed framework is examined on two different HAR data sets demonstrating a possibility of data-driven machine learning for near an optimal classification of activities. The proposed framework exhibited effectiveness and efficiency when compared with the existing methods. The modified feature engineering resulted in 42% fewer features required by support vector machine to yield 97.3% correct recognition of human physical activities. However, the addition of post-learning data engineering further improved the model to perform 99% accurate classification, which is an almost optimal performance.

In the field of pattern recognition and data learning process the dimensionality reduction is a necessary method for providing a personalized data in perfect numbers. The increase in research proclaims the importance of data reduction in current trends. The old traditional reduction techniques use ranking systems. But this differs from others in reduction of data without lagging its performance in terms of increased efficiency and less percentage of error occurrences. The area of dimensionality reduction not only ends in pattern recognition and also in many high dimensional data processing elements such as text categorization, indexing of documents and mainly in gene expression data. The feature extraction and feature selection are the two steps of process in reduction. The process proposed is statistical pattern recognition process and a paradigm for this approach to the problem is summarized herein. The four data process includes is (1) evaluation (2) acquisition (3) feature selection, and (4) statistical model of feature selection. We show how to reduce the dimensionality by utilizing group and convexity model. This paper concluded by an integrated approach to the implementation dimensionality reduction process with an effort to maximize the efficiency and accuracy.

Facial sentiment analysis has been an enthusiastic research area for the last two decades. A fair amount of work has been done by researchers in this field due to its utility in numerous applications such as facial expression driven knowledge discovery. However, developing an accurate and efficient facial expression recognition system is still a challenging problem. Although many efficient recognition systems have been introduced in the past, the recognition rate is not satisfactory in general due to inherent limitations including light, pose variations, noise, and occlusion. In this paper, a hybrid approach of facial expression based sentiment analysis has been presented combining local and global features. Feature extraction is performed fusing the histogram of oriented gradients (HOG) descriptor with the uniform local ternary pattern (U-LTP) descriptor. These features are extracted from the entire face image rather than from individual components of faces like eyes, nose, and mouth. The most suitable set of HOG parameters are selected after analyzing them experimentally along with the ULTP descriptor, boosting performance of the proposed technique over face images containing noise and occlusions. Face sentiments are analyzed classifying them into seven universal emotional expressions: Happy, Angry, Fear, Disgust, Sad, Surprise, and Neutral. Extracted features via HOG and ULTP are fused into a single feature vector and this feature vector is fed into a Multi-class Support Vector Machine classifier for emotion classification. Three types of experiments are conducted over three public facial image databases including JAFFE, MMI, and CK+ to evaluate the recognition rate of the proposed technique during experimental evaluation; recognition accuracy in percent, i.e., 95.71, 98.20, and 99.68 are achieved for JAFFE, MMI, and CK+, respectively.

Face recognition is a challenging research field in computer sciences, numerous studies have been proposed by many researchers. However, there have been no effective solutions reported for full illumination variation of face images in the facial recognition research field. In this paper, we propose a methodology to solve the problem of full illumination variation by the combination of histogram equalization (HE) and Gaussian low-pass filter (GLPF). In order to process illumination normalization, feature extraction is applied with consideration of both Gabor wavelet and principal component analysis methods. Next, a Support Vector Machine classifier is used for face classification. In the experiments, illustration performance was compared with our proposed approach and the conventional approaches with three different kinds of face databases. Experimental results show that our proposed illumination normalization approach (HE_GLPF) performs better than the conventional illumination normalization approaches, in face images with the full illumination variation problem.

In the image handling and image processing areas, most operations can be executed in a pixel-by-pixel or cluster-by-cluster manner. These parallel and simultaneous executions have many benefits, and many researchers showed remarkable improvements. In this paper, we started from a specific and practical image handling and feature extraction sequences. We focused on the detailed design and robust implementation on the modern massively parallel architecture of CUDA. We present the enhanced features of our implementation and their design details. Our final result shows 13 times faster execution speed, in comparison with its previous CPU-based implementation. These methods can be applied to the variety of image manipulation processes.

Various types of three-dimensional (3D) cameras have been used to analyze real-world objects or environments effectively. However, because most existing 3D cameras capture scenes by statically using one type of camera, there may be a limit to the quality of the captured images. Therefore, in this paper, we build a hybrid camera system that combines passive triangulation (PT)- and active triangulation (AT)-based cameras and suggest a new mechanism of estimating accurate 3D depth by adaptively switching between the two types of cameras depending on the complexity of the environment. The suggested method initially uses initial input images to extract brightness and texture, which are major features representing the current state of the surrounding environment. The method subsequently generates a set of rules that dynamically select the PT- or AT-based camera, whichever can operate more suitably in the current environment, by analyzing the two extracted features. In experimental results, we demonstrate that the proposed adaptive camera-selection approach can be applied to extract 3D depth reliably with reasonable performance in terms of accuracy and time.