Search

In the machine learning jargon, multi-label classification refers to a task where multiple mutually non-exclusive class labels are assigned to a single instance. Generally, the lack of sufficient labeled training data demanded by a classification task is met by an approach known as semi-supervised learning. This type of learning extracts the decision rules of classification by utilizing both labeled and unlabeled data. Regarding multi-label data, however, current semi-supervised learning methods are unable to classify them accurately. Therefore, with the goal of generalizing the state-of-the-art semi-supervised approaches to multi-label data, this paper proposes a novel two-stage method for multi-label semi-supervised classification. The first stage determines the label(s) of the unlabeled training data by means of a smooth graph constructed using the manifold regularization. In the second stage, thanks to the capability of the twin support vector machine to relax the requirement that hyperplanes should be parallel in classical SVM, we employ it to establish a multi-label classifier called LP-MLTSVM. In the experiments, this classifier is applied on benchmark datasets. The simulation results substantiate that compared to the existing multi-label classification algorithms, LP-MLTSVM shows superior performance in terms of the Hamming loss, average precision, coverage, ranking loss, and one-error metrics.

A linear controller for the unstable flywheel inverted pendulum (also known as a reaction wheel inverted pendulum or inertia wheel pendulum) is designed using Quantitative Feedback Theory (QFT) to provide response that stays withing upper and lower bounds in the face of significant variations of the plant parameters. A mathematical model of the plant, including both the mechanics of the pendulum and the electromechanics of the motor, is derived. Simulation is used to compare performance of the QFT controller to that of simple PID compensation. It is shown that while the QFT controller consistently satisfies the robust performance bounds for all plant uncertainty in both frequency and time domains, the PID controller fails to provide robust tracking performance.

In this paper, a proposed structure of a parallel hybrid power filter is introduced which is appropriate for group elimination of current harmonics as well as reactive power compensation in medium or low voltage power grids. According to the common structures of parallel hybrid filters, the proposed structures include a series resonance circuit with a small inductance in parallel with the active power filter. In other words, the passive filter design is a hybrid structure in which the majority of the main reactive current passes through additional inductance instead of passing through the active power filter, which reduces the current flowing through the switching equipment of the power converter. Also, due to the special design of the passive filter in this proposed structure, the active power filter does not need to withstand the harmonic voltage. Therefore, the result and advantage of these features is that the volt-ampere of the active power filter is significantly reduced. Computer simulation results indicate strong application of the proposed parallel HAPF to harmonics exclusion and reactive power compensation.

There has been tremendous research progress in estimating the depth of a scene from a monocular camera image. Existing methods for single-image depth prediction are exclusively based on deep neural networks, and their training can be unsupervised using stereo image pairs, supervised using LiDAR point clouds, or semi-supervised using both stereo and LiDAR. In general, semi-supervised training is preferred as it does not suffer from the weaknesses of either supervised training, resulting from the difference in the cameras and the LiDARs field of view, or unsupervised training, resulting from the poor depth accuracy that can be recovered from a stereo pair. In this paper, we present our research in single image depth prediction using semi-supervised training that outperforms the state-of-the-art. We achieve this through a loss function that explicitly exploits left-right consistency in a stereo reconstruction, which has not been adopted in previous semi-supervised training. In addition, we describe the correct use of ground truth depth derived from LiDAR that can significantly reduce prediction error. The performance of our depth prediction model is evaluated on popular datasets, and the importance of each aspect of our semi-supervised training approach is demonstrated through experimental results. Our deep neural network model has been made publicly available., Comment: Submitted to IROS2019

Video stabilization and mosaicking is an important task when a video stream of a large area, such as the video stream from a UAV or a blimp, is received. In this paper we propose a novel approach for video stabilization and mosaicking from shaky video streams. We present a feature-based real-time video mosaicking pipeline performing image alignment by combining feature point detector, descriptor, a robust statistical selection methodology (RANSAC), and other filters. By using this method, we can eliminate the mechanical stabilizer. The stabilization is done using SURF feature detectors, BRSIK descriptor, and affine transition matrix. To reach real-time performance, we used C++ OpenCV libraries for implementation and compared most feature detectors and descriptors by their time consumption factor to find out which one is better for real-time approach. Finally, the approach has been tested on real video streams whic showed good performance.

Build classifier based on fuzzy rules for high-dimensional data sets, such as genetic data, are faced with great difficulties. An effective approach to this problem using feature selection techniques and dimension reduction methods. Hence, in this paper, using five different feature selection methods, size of data is reduced and the based on accuracy of the support vector machines classifier to this data a five dimensional feature vector extracted then using frog leaping algorithm and genetic algorithm, With the aim of minimizing the number of rules and optimize the parameters of its a set of fuzzy rules for data classification are extracted. The proposed method was tested on five gene expression datasets. The experiments results show that the proposed method achieves higher accuracy than existing.

This paper presents an electrowetting-on-dielectric (EWOD)-based droplet mixing and splitting device. In the proposed method, the droplet is sandwiched between the bottom and top substrate plates of the EWOD device. A dielectric layer of Parylene C and a thin coating of a hydrophobic Teflon layer are used to create the EWOD device. The device actuates the droplets by applying an electric potential and thus increases the wettability of the droplet on the EWOD surface. The Finite Element Method (FEM) based Coventorware software is used to accomplish the simulations. The simulated result shows that a significant change in contact angle (120° to 80°) of the droplet is occurred during the operations of both mixing and splitting. The EWOD device with Parylene C layer of 2 μm thickness and Teflon layer of 50 nm thickness started both droplet mixing and splitting operations at a low actuation voltage of 30 V.

A data stream is an ordered and continuous sequence of examples that can be examined only once. Data stream mining introduces new challenges compared to traditional mining algorithms. Fuzzy c-means (FCM) is a method of clustering in which a data point can assign to more than one cluster at the same time. In this paper we extend FCM algorithm to clustering data streams. Our performance experiments over KDD-CUP'99 data set show the efficiency of the algorithm.

Anthropogenic emissions of CO 2 account for one third of the entire CO 2 emissions across the world. Cement industry is considered to be one of the major sources of industrial CO 2 emissions which are about 1.8 Gt/y. These emissions are produced both from fuel combustion and cement production process. Coal as one the major fuels used widely by US cement industry, is considered to be a major source of CO 2 emissions. CO 2 emissions reduction can be achieved by either choosing an appropriate fuel processing technologies or capturing it post-processing from flue gas. This paper is a brief review of CO 2 emissions and their associated sources in the US cement industry. Conventional and modern approaches of increasing energy efficiency in the cement production process as well as CO 2 capture methodologies were studied. Effect of CCS (Carbon Capture and Sequestration) methods integration to the cement plant costs was also reviewed.

One of the important risk factors in pervasive computing environments is uncertainty. The cause of this uncertainty, changes in user mobility or transport user. In general, uncertainty in pervasive computing environments occurs in two cases: uncertainty in sensors and uncertainty in sensor information combination. Error in the verification and misunderstanding can be caused to fall risk pervasive computing applications security, power, efficiency and performance. In fact, access control in this environment is dependent on the evaluation of sensors and information received from them. In other words, if the information is correct, that is received through the sensors, can have access control with high reliability.

This paper illustrates the use of Mixture of Experts technique to solve low resolution faces recognition problem. Proposed method is based on extracting diverse features from low resolution images using three well known feature extractors; PCA, DCT and FFT, and applying ME architecture on the output vector of each feature extractor. The experimental results prove our claims that this method is superior to many similar methods such as Stacked Generalization-Based methods. By the way, this method is faster than Stacking. The enhancement in performance of the proposed method than Stacked Generalization was 1% on the same database.

One of the most significant parameter in increasing the efficiency of MLP NN that utilizes the EBP algorithm for training network is convergence speed which different methods have been proposed for improving it. In this paper, we use a variable learning rate method for increasing the convergence speed of EBP algorithm, which its idea have come from a one way presented to improve the efficiency of Standard LMS. The result of comparison of standard EBP and proposed VSSEBP algorithm over various datasets demonstrate that VSSEBP have high convergence speed. All experiments have performed on noisy data with various SNR values.

This paper presents an appropriate solution for low resolution faces recognition problem, using combination of diverse classifiers. We investigate our model based on extracting important features from low resolution images using three well known feature extractors; PCA, DCT and FFT, assigning MLP classifiers to each feature extractor and combining the votes of MLP classifiers using fusion of experts techniques. The results show that using the combination of three mentioned feature extractors and applying the Stack Generalization as combiner of classifiers for low resolution face recognition task, leads to higher performance than other recognition models. It's worth nothing that using a single feature extractor for low resolution face recognition task was failed in the previously surveys.

This paper deals with Single training sample Face recognition which is a new challenging problem in pattern recognition and machine vision. In the proposed model, Gabor filter is applied on face images, to generate four Gabor responses. And then, the generated samples along with the original image used as the input of five Nearest Neighbor Base-classifiers. To increase the correct recognition rate, the votes of Base-Classifiers are combined using majority voting technique. The correct class of each test sample is determined by the majority of the classifiers votes. But in cases which the winner class can't be determined using majority voting (deep uncertainty), we use Enhanced Majority voting which chooses the vote of original image's Base-Classifier as the winner class. The resultant recognition rate of applying this model on ORL face database was improvement in recognition rate about 2%, 4% and 5% than 2DPCA, (PC)2 A and PCA respectively.

In classifiers combination, the diversity rate among classifier's outputs is one of the most important discussions. There are different methods for combining classifiers. AdaBoost is an incremental method for creating a classifiers ensemble in which every AdaBoost algorithm has a local centrality. It means that classifiers are data biased and classify special data. In this paper we intend to find a new method for combining classifiers by using AdaBoost method and diversity concept. We have checked this method over different data sets and compared results of this method with others. These results indicate that we can develop other versions of this method for achieving a better performance.