Intelligent Computing Methodologies

This paper proposes an adaptive position synchronization controller using orthogonal neural network for 3-DOF planar parallel manipulators. The controller is designed based on the combination of computed torque method with position synchronization technique and orthogonal neural network. By using the orthogonal neural network with online turning gains can overcome the drawbacks of the traditional feedforward neural network such as initial values of weights, number of processing elements, slow convergence speed and the difficulty of choosing learning rate. To evaluate the effectiveness of the proposed control strategy, simulations were conducted by using the combination of SimMechanics and Solidworks. The tracking control results of the parallel manipulators were significantly improved in comparison with the performance when applying non-synchronization controllers.

One of the important problems of robotics is the navigation of mobile robots in uncertain environments that are densely cluttered with obstacles. The control of robots using the traditional control algorithms is not satisfactory as far as the navigational accuracy and the distance and time to reach the goal are concerned, when the robot is in a complicated surrounding. One of alternative and efficient ways of constructing a control system that explicitly deal with uncertainty is the use of fuzzy systems approach. The paper is devoted to navigation of mobile robot using type-2 fuzzy system. The design principle of navigation algorithm using type-2 fuzzy system is presented. The fuzzy knowledge base that describes the relation between the input- current angle and distance signals and output signals that determine the robot turn angle is developed. The control rules of navigation and inference engine operations have been described. The comparative simulation results of robot navigation system demonstrate the advantage of the fuzzy navigation algorithm.

The performance of non-rigid structure from motion (NRSFM) generally deteriorates when the image sequence is small. In this paper, an effective approach is proposed to deal with NRSFM with small size sequences based on the Expectation and Maximization-Procrustean Normal Distribution (EM-PND) algorithm. In the proposed method, the sub-sequences are first extracted from the original small size sequence. Further, some weaker estimators are constructed by inputting the sub-sequences to the EM-PND algorithm. Finally, the 3Dstructures of the sequences are estimated by integrating the outputs of these weaker estimators. Experimental results on several widely used sequences demonstrate the effectiveness and feasibility of the proposed algorithm.

Research interest in robotic wheelchairs is driven in part by their potential for improving the independence and quality-of-life of persons with disabilities and the elderly. Moreover, smart wheelchair systems aim to reduce the workload of the caregiver. In this paper, we propose a novel technique for 3D sensing of the terrain using a conventional Laser Range Finder (LRF). We mounted this sensing system onto our new six-wheeled robotic step-climbing wheelchair and propose a new step measurement technique using the histogram distribution of the laser data. We successfully measure the height of stair steps in a railway station. Our step measurement technique for the wheelchair also enables the wheelchair to autonomously board a bus. Our experiments show the effectiveness and its applicability to real world robotic wheelchair navigation.

Recognizing human emotion by computer vision is an interesting and challenging problem. In particular, the reading of inner emotions, has received limited attention. In this paper, we use a remote video-based heart rate sensing technique to obtain physiological data that provides an indication of a person’s inner emotions. This method allows for contactless estimates of heart rate data while the subject is watching emotionally stimulating video clips. We also compare against a wearable heart rate sensor to validate the usefulness of the proposed remote heart rate reading framework. We found that the reading of heart rates of a subject effectively detects the inner emotional reactions of human subjects while they were watching funny and horror videos—despite little to no facial expressions at times. These findings are validated from the reading of heart rates for 40 subjects with our vision-based method compared against conventional wearable sensors. We also find that the change in heart rate along with emotionally stimulating content is statistically significant and our remote sensor is well correlated with the wearable contact sensor.

Agricultural pests can cause serious damage to crops and need to be identified during the agricultural pest prevention and control process. In comparison with the low-speed and inefficient artificial identification method, it is important to develop a fast and reliable method for identifying agricultural pests based on computer vision. Aiming at the problem of agricultural pest identification in complex farmland environment, a recognition method through deep learning is proposed. The method could recognize and track the agricultural pests in surveillance videos of farmlands by using deep convolutional neural network and Faster R-CNN models. Compared with the traditional machine learning methods, this method has higher recognition accuracy in high background noise, and it can still effectively recognize agricultural pests with protective colorations. Therefore, compared with the current agricultural pest static-image recognition method, this method has a higher practical value and can be put into the actual agricultural production environment with the agricultural networking technology.

Face detection is the first step for many facial analysis applications and has been extensively researched in the visible spectrum. While significant progress has been made in the field of face detection in the visible spectrum, the performance of current face detection methods in the thermal infrared spectrum is far from perfect and unable to cope with real-time applications. As the Viola-Jones algorithm has become a common method of face detection, this paper aims to improve the performance of the Viola-Jones algorithm in the thermal spectrum for detecting faces with or without eyeglasses. A performance comparison has been made of three different features, HOG, LBP, and Haar-like, to find the most suitable one for face detection from thermal images. Additionally, to accelerate the detection speed, a pre-processing stage is added in both training and detecting phases. Two pre-processing methods have been tested and compared, together with the three features. It is found that the proposed process for performance enhancement gave higher detection accuracy (95%) than the Viola-Jones method (90%) and doubled the detection speed as well.

In most of classic plant identification methods a dichotomous or multi-access key is used to compare characteristics of leaves. Some questions about if the analyzed leaves are lobed, unlobed, simple or compound need to be answered to identify plants successfully. However, very little attention has been paid to make an automatic distinction of leaves using such features. In this paper we first explore if incorporating prior knowledge about leaves (categorizing between lobed simple leaves, and the unlobed simple ones) has an effect on the performance of six classification methods. According to the results of experiments with more than 1,900 images of leaves from Flavia data set, we found that it is statically significant the relationship between such categorization and the improvement of the performances of the classifiers tested. Therefore, we propose two novel methods to automatically differentiate between lobed simple leaves, and the unlobed simple ones. The proposals are invariant to rotation, and achieve correct prediction rates greater than 98%.

Pedestrian tracking has been arguably addressed as a special topic beyond general object tracking. Although many learning or data driven object trackers as well as recent deep learning object trackers have shown excellent performance for general object tracking, they have limited success on pedestrian tracking because there exist three major challenges emerging from pedestrian tracking such as vast variations of human bodies, distraction from similar persons and complete occlusion. In this paper, we propose an integrated learning framework for pedestrian tracking to overcome these problems. It is demonstrated by the experimental results on the SVD-B dataset that our proposed framework can achieve competitive results in comparison with state-of-the-art object trackers under the evaluation of the precision and success rate as well as fps.

Disc herniation is a major reason for lower back pain (LBP), a health issue that affects a very high proportion of the UK population and is costing the UK government over £1.3 million per day in health care cost. Currently, the process to diagnose the cause of LBP involves examining a large number of Magnetic Resonance Images (MRI) but this process is both expensive in terms time and effort. Automatic labeling of lumbar disc pixels in the MRI to detect the herniation area will reduce the time to diagnose and detect the cause of LBP by the physicians. In this paper, we present a method for automatic labeling of the lumbar spine disc pixels in axial view MRI using pixels locations and gray level as features. Clinical MRIs are used for the training and testing of the method. The pixel classification accuracy and the quality of the reconstructed disc images are used as the main performance indicators for our method. Our experiments show that high level of classification accuracy of 91.1% and 98.9% can be achieved using Weighted KNN and Fine Gaussian SVM classifiers respectively.

Researches show in the test phase of Convolutional Neural Network (CNN), the most time cost occurred in convolutional layers, while the most memory cost occurred in the fully connected layers. With the rapid development of the pedestrian methods, which are based on deep learning, the performance is going better and better. Especially using resemble models, pedestrian detection can get a more excellent performance. However, the performance is improved by the increase in parameters and slow of speed in price. Meanwhile, for some specific tasks, such as driverless cars, due to the limitations of hardware facilities, it is impossible to use these methods on them. In this paper, we applied the tensor decomposition to pedestrian detection task, in order to accelerate the whole pedestrian detection processing. Experiments show even though the decomposition brings some of the rise of miss rate (MR), the saved memory and time indicates the efficiency of our method.

Popular registration methods can be applying into multimodal images, such as Harris-PIIFD, SURF-RPM, GMM, GDB-ICP and so on. There exist some challenges in existing multimodal image registration techniques: (1) They fail to register image pairs with some significantly different content, illumination and texture changes; (2) They fail to register image pairs with too small overlapping or too much noise. To address these problem, this paper improves the multimodal registration by contribute a novel robust framework SURF-PIIFD-BBF-VFC (SPBV). The SURF-PIIFD method can provide enough repeatable and reliable local features; the bilateral matching method and vector field consensus (VFC) can establish robust point correspondences of two point sets. For evaluation, we compare the performance of the proposed SPBV with two existing methods Harris-PIIFD and SURF-RPM on two multimodal data sets. The results indicate that our SPBV method outperforms the existing methods and it is robust to low quality and small overlapping multimodal images.

This paper presents a sectional algorithm for indoor location using wireless sensor networks. This algorithm uses the motion regularity of target to compute the next motion area quickly and apply the pre-processed compressive sensing method to that area, which reduce the location problem to a sparse signal reconstruction problem. Then we carry out the proposed algorithm on the motion of next time turn by turn, such procedure is able to locate with fewer data collection, wireless links and wireless nodes as well as raise the accuracy of location. The simulation results show that the proposed algorithm of dynamic motion based compressive sensing sectional location method has a good performance.

Crowdsourcing is used to leverage external crowds to perform specialized tasks quickly and inexpensively. In the application of crowdsourcing, one task may often include many ordered steps. Based on the different requirements (e.g. workers’ skills and etc.) of steps, the task requester may divide the task into many sub-tasks, and publish the subtasks in the crowdsourcing system. Moreover, service requesters usually want to finish their submitted tasks as immediately as possible. However, there has been no allocation mechanism with consideration of crowdsourcing tasks with precedence constraints and minimization of the total execution time simultaneously. To tackle this challenge, we consider the precedence constraints among tasks and design an efficient task allocation mechanism for the crowdsourcing system. In this work, we first introduce the crowdsourcing system model and formulate the task allocation problem. After proving that the studied problem is NP-hard, we propose an approximation algorithm that can minimize the total execution time of all the tasks. Then, we conduct extensive simulations to evaluate the performance of the proposed algorithm, and the simulation results show that the proposed algorithm has good approximate optimal ratios under different parameter settings.

This paper proposes a link layer topology discovery algorithm based on dynamic programming. The basic idea of this algorithm is to express the topology of the switch and the switch, the switch and the host in the form of tree. According to the principle of multi-stage decision process, we set to construct a single order tree for the whole network topology using the address forwarding table (AFT) port. The results of theoretical analysis and practical application show that the topology discovery algorithm has been greatly improved in terms of efficiency, accuracy and effectiveness.

This paper abstracts a practical problem from the iron and steel enterprise, and researches a multiple crane integrated scheduling problem with flexible group decision (crane grouping scheduling for short). For this demonstrated NP-hard problem, we propose a heuristic algorithm based on some analyzed properties. For a restrict case, we analyze the worst-case performance. Further, for the general case, the average performance of the heuristic algorithm is computationally evaluated. The results show that the proposed heuristic algorithm is capable of generating good quality solutions.

In this paper, a novel approach has been proposed for fault diagnosis of internal combustion (IC) engine using Empirical Mode Decomposition (EMD) and Neural Network. Live signals from the engines were collected with and without faults by using four sensors. The vibration signals measured from the large number of faulty engines were decomposed into a number of Intrinsic Mode Functions (IMFs). Each IMF corresponds to a specific range of the frequency component embedded in the vibration signal. This paper proposes the use of EMD technique for finding IMFs. The Cumulative Mode Function (CMF) was chosen rather than IMFs since all the IMFs are not useful to reveal the vibration signal characteristics due to the effect of noise. Statistical parameters like shape factor, crest factor etc. of the envelope spectrum of CMF were investigated as an indicator for the presence of faults. These statistical parameters are used in turn for classification of faults using Neural Networks. Resilient Propagation which is a rapidly converging neural network algorithm is used for classification of faults. The accuracy obtained by using EMD-ANN technique effectively in IC engine diagnosis for various faults is more than 85% with each sensor. By using a majority voting approach 96% accuracy has been achieved in fault classification.

This paper details a robust collaborative intrusion detection methodology for detecting attacks within a Cloud federation. It is a proactive model and the responsibility for managing the elements of the Cloud is distributed among several monitoring nodes. Since there are a wide range of elements to manage, complexity grows proportionally with the size of the Cloud, so a suitable communication and monitoring hierarchy is adopted. Our architecture consists of four major entities: the Cloud Broker, the monitoring nodes, the local coordinator (Super Nodes), and the global coordinator (Command and Control server - C2). Utilising monitoring nodes into our architecture enhances the performance and response time, yet achieves higher accuracy and a broader spectrum of protection. For collaborative intrusion detection, we use the Dempster Shafer theory of evidence via the role of the Cloud Broker. Dempster Shafer executes as a main fusion node, with the role to collect and fuse the information provided by the monitors, taking the final decision regarding a possible attack.

With the increasing of website user experiences, the interactive UI effect adds the complexity of web applications, which results in the development of Web front-end frameworks. Nevertheless, the existing code quality metrics models are not able to measure the code quality of Web front-end frameworks effectively. On the other hand, it is also difficult to present a quality metrics model that can evaluate the software frameworks comprehensively, so we should focus on a specific measurement object in order to improve accuracy of metrics. On this groundwork, we present a code quality metrics model for Web front-end frameworks based on React. In this model, we put forward 16 metric units, according to the characteristics of JavaScript and React through experiments. Then we determine the quantitative rules for each metric unit. At long last, we generate a set of scores to measure Web applications. In the part of experiment, firstly we select two projects as our benchmarks, then we choose a real project to evaluate the code qualify and compare the results with people’s evaluation of this project. The experiment results imply that this model is reliable and has some directing significance. The primary contributions of this paper are proposing a new code quality metrics model for JavaScript and React, which is a new research direction in the area of code quality metrics. By this model, we can monitor code quality and enhance development efficiency.

Internet of Things (IoT) becomes a reality for many reasons: low powers processors, improvements in wireless communication technologies and electronic devices. This will initiate new business opportunities in providing these novel applications and services, which integrate efficiently IoT services into enterprise applications. The work presented here proposes an Agent-based approach for an effective integration of the IoT in enterprise systems. The proposed approach defines a new meta-model to describe this integration. Also, the paper presents an agent-based system architecture whose main goal is to address and tackle interoperability challenges in the context of IoT environments. Also, it solves the interoperability issues between heterogeneous Cloud services environments by offering a harmonized API.

In this paper, an intelligent control scheme based on fuzzy proportional-integral-derivative controller (FPIDC) is proposed for distributed generations in islanded microgrid. The proposed FPIDC method is composed of a close-loop of the virtual impedance to compensate the feeder mismatch between distributed generators (DGs). Together with feedback of the inaccurate reactive power, the uncertainty of proportional-integral-derivative (PID) parameters is removed by adaptive tuning. Therefore, the power sharing performance is improved and the circulating current between DGs is mitigated. The dynamic response of the microgrid system with the proposed FPIDC is much better than that of the conventional PID controller. The comparison and analysis of the proposed control with conventional control are carried out to evaluate the superiority of the proposed method.

DC-DC converter is widely used in industrial applications with the aid of proportional-integral-derivative (PID) controller. The PID controller is commonly used because of its simple, efficiency and small steady state error. However, DC-DC converter is mainly nonlinear system due to its inherent switching operation, which makes the PID parameters very difficult to be found. In this paper, an intelligent PID controller based on a hybrid combination with the fuzzy logic controller (FLC) is proposed to control DC-DC converter. The proposed fuzzy based PID controller (FPIDC) is a close-loop control of the nonlinear PID parameters by measuring output voltage of the DC-DC converter, and it is designed effectively to reduce computation burden. By tuning the PID parameters continuously according to the load condition, the performance of the DC-DC converter is improved much better. The comparison between the proposed and conventional PID control method are investigated under various operating conditions. Simulation results and model analysis are carried in MATLAB/SIMULINK to prove the effectiveness of the proposed FPIDC method.

This paper presents a method for random variable generation based on $$ \alpha $$-cuts. The proposed method uses convex fuzzy numbers with single-element core and uniformly distributed random numbers to obtain random variables, mainly used in simulation models.

In this paper, a heuristic search technology is presented for the path planning of a bionic robotic fish equipped with single beam sonar. When the bionic robotic fish is navigating, the scanning range of the sonar is divided into 5 parts, and each sub-part is a candidate course, while the middle part as the current course. With the fuzzy relation between sub-parts of sonar and real time environment attributes as the core concept, the triangle sub-product relationship put forward by Bandler and Kohout reveals the relationship between the various parts of the sonar, and a sub-part of sonar is selected as the inheritance course of bionic fish. The simulation scene and the pool scene are set to validate the effect of the planning strategy, as a result of which it can conclude that the planning strategy based on BK triangle sub-product can effectively realize path planning of bionic robotic fish.

This paper studies the problem of scheduling customers or tasks in a queuing system. Generally the customers or a set of tasks in queuing system are attended according with different rules as round robin, equiprobable, shortest queue, among others. However, the condition of the system like the work in process, utilization and the length of queue is difficult to measure. We propose to use a fuzzy inference system in order to determine the status in the system depended of input variables like the length queue and the utilization. The experiment results shows an improvement in the performance measures compared with traditional scheduling policies.

Based on the analysis of the main influencing factors of campus network information security, the risk evaluation index system of campus network information security is constructed. Meanwhile, a security risk assessment model of campus network information security based on FAHP and matter element model is proposed. It is used to complete the quantification of index, the determination of index weight and the calculation of correlation degree. The model can also be used to complete the campus network information security risk assessment, and then put forward the improvement measures. The results show that the model can be effectively applied to the campus network information security risk assessment. It provides a theoretical basis for improving campus network information security.

A novel semi-supervised classification algorithm for short text based on fusion similarity is presented via analyzing of existing defects of short text classification algorithm. First of all, some words with the ability of indication of the category are extracted from the labeled dataset to construct a strong category features set. A valid fusion similarity measurement method is designed by combining cosine theorem and strong category features based similarity. Secondly, computing the mean value of the supervised information, and determining the virtual class center point of each class, and then finding the real class center point. Finally, we search those texts which have the highest similarity with each real class center in the unlabeled dataset, and give it the same class label with the real class center point. At the same time, we add it to the labeled collection, update the strong category features set and the similarity matrix. Repeat this process until all short texts have been labeled. Ultimately, experiments show that our method can significantly improve the efficiency of short text classification. The text of the most similarity with the center of the class.

High dimensional data clustering is a difficult task due to the curse of dimensionality. Traditional clustering methods usually fail to produce meaningful results for high dimensional data. Hypergraph partition is believed to be a promising method for dealing with this challenge. In this work, a new high dimensional clustering method called Merging Dense SubGraphs (MDSG) is proposed. A graph G is first constructed from the data by defining an adjacency relationship between the data points using Shared k Nearest Neighbors (SNN). Then a hypergraph is created from the graph G by defining the hyperedges to be all the maximal cliques in the graph. After the hypergraph is produced, an improved hypergraph partitioning method is used to produce the final clustering results. The proposed MDSG method is evaluated on several real high dimensional handwritten datasets, and the experimental results show that the proposed method is superior to the traditional clustering method and other hypergraph partition methods for high dimensional handwritten data clustering.

K-means is the most well-known and widely used classical clustering method, benefited from its efficiency and ease of implementation. But k-means has three main drawbacks: the selection of its initial cluster centers can greatly affect its final results, the number of clusters has to be predefined, and it can only find clusters of similar sizes. A lot of work has been done on improving the selection of the initial cluster centers and on determining the number of clusters. However, very little work has been done on improving k-means to deal with clusters of different sizes. In this paper, we have proposed a new clustering method, called k-normal, whose main idea is to learn cluster sizes during the same process of learning cluster centers. The proposed k-normal method can identify clusters of different sizes while keeping the efficiency of k-means. Although the Expectation Maximization (EM) method based on Gaussian mixture models can also identify the clusters of different sizes, it has a much higher computational complexity than both k-normal and k-means. Experiments on a synthetic dataset and seven real datasets show that, k-normal can outperform k-means on all the datasets. If compared with the EM method, k-normal still produces better results on six out of the eight datasets while enjoys a much higher efficiency.

Kernel based Non-Negative Matrix Factorizations (KNMFs) are one of the most important methods for non-negative nonlinear feature extractions and have achieved good performance in pattern classifications. However, most existing KNMF algorithms are merely valid for one special kernel function. Also, they model the pre-images inaccurately. In this paper, we utilize kernel matrix learning strategy to develop a Universal KNMF (UKNMF) algorithm, which is able to use all Mercer kernel functions. The proposed method avoids the pre-image learning simultaneously. We first establish three objective functions and then derive three update formula to determine three matrices, namely one feature matrix and two kernel matrices. The iterative rules are theoretically proven to be convergence by means of auxiliary function technique. Our UKNMF approaches with polynomial kernel and RBF kernel (UKNMF-Poly and UKNMF-RBF) are applied to face recognition respectively. The face databases, including ORL and Yale face databases, are selected for evaluations. Compared with some state of the art kernel based algorithms, experimental results show the effectiveness and superior performance of the proposed methods.

Membrane protein is a kind of protein with unique transmembrane structure, which is the material basis for cells to perform various functions. It is an important biological signal molecule to assume the information transmission between the cell and the external environment. It is a precursor step to predict the classification of β-barrel transmembrane protein according to the protein sequence information for 3D structure modeling and function analysis. We firstly use the method of compromising features consist of the position information in sequence and the physiochemical properties of amino acid residues. Then a model by support vector machine algorithm (SVM) is built to predict the β-barrel transmembrane protein. The experimental results presented that transmembrane protein structure prediction based on SVM can provide valid enhancement to transmembrane protein 3D structure prediction and function analysis.

Nonlinear problems encountered in theoretical study on the financial risks are important objects of nonlinear science. Traditional financial risk management theory of nonlinear problems unresolved can be studied by using the method of nonlinear science, financial soliton theory and big data ideation. The theory can not only analyze the evolution of financial markets, the formation of risk transfer mechanism, but also it can more profoundly grasp the behaviors of financial markets, obtain new financial risk prediction, control concepts and methods.

Classification of imbalanced datasets is one of the challenges in machine learning and data mining domains. The traditional classifiers still need to handle with minority instances. In this paper, we propose an effective method which applies sampling method based on ensemble learning. It uses Adaboost-SVM based on spectral clustering to boost the performance. This method also uses over-sampling and under-sampling methods based on the misclassified instances got by ensemble learning. Compared with the preview algorithms, the experiment results show that the proposed method is effective in dealing with imbalanced data in binary classification.

The commonly used algorithms in recommender system tend to recommend popular items. The recently proposed algorithm, denoted as G-CosRA, shows good performance in handling this problem, with two parameters to control the popularity of items and activeness of users. In this paper, we refine this algorithm and propose a new recommendation algorithm based on adaptive Kendall’s τ correlation, where only one tuning parameter is involved. The proposal has better performance in accuracy, popularity and diversity, compared with G-CosRA and other existing algorithms. A parameter-free version, named weighted Kendall, is also proposed for better efficiency in computing.

Geospatial clustering is an important topic in spatial analysis and knowledge discovery research. However, most existing clustering methods clusters geospatial data at data level without considering domain knowledge and users’ goals during the clustering process. In this paper, we propose an ontology-based geospatial cluster ensemble approach to produce better clustering results with the consideration of domain knowledge and users’ goals. The approach includes two components: an ontology-based expert system and a cluster ensemble method. The ontology-based expert system is to represent geospatial and clustering domain knowledge and to identify the appropriate clustering components (e.g., geospatial datasets, attributes of the datasets and clustering methods) based on a specific application requirement. The cluster ensemble is to combine a diverse set of clustering results which is produced by recommended clustering components into an optimal clustering result. A real case study has been conducted to demonstrate the efficiency and practicality of the approach.

Nowadays, people are increasingly inclined to use social tools to express their intentions explicitly and implicitly. Most of the work is dedicated to solving the explicit intention detection, ignoring the implicit intention detection, as the former is relatively easy to solve with the classification method. In this work, we use the Attention-Based Encoder-Decoder model which is specified for the sequence-to-sequence task for user implicit intention detection. Our key idea is to leverage the model to “translate” the implicit intention into the corresponding explicit intent by using the parallel corpora built on the social data. Specifically, our model has domain adaptability since the way people express implicit intentions for different domain is variable, while the way to express explicit intentions is mostly in the same form, such as “I want to do sth”. In order to demonstrate the effectiveness of our method, we conduct experiments in four domains. The results show that our method offers a powerful “translation” for the implicit intentions and consequently identifies them.

Information explosion causes a serious scarcity of people’s time and a severe divergence of people’s attention. This paper addresses the issue of automatic summarization for Korean texts and presents a novel keyword-extraction-based Korean text summarization (KKTS) algorithm. We investigate the enhancement of POS-tagging to the KKTS algorithm according to three kinds of text feature: noun words, predicate words, and all words. The experimental results show that our POS-tagging enhanced KKTS algorithm according to noun words can achieve the best performance in the Korean summarization task.

Most supervised term weighting (STW) schemes can only be applied to binary text classification tasks such as sentiment analysis (SA) rather than text classification with more than two categories. In this paper, we proposed a new supervised term weighting scheme for multi-class text categorization. The so-called inverse term entropy (ite) measures the distribution of different terms across all the categories according to the definition of entropy in information theory. We present experimental results obtained on the 20NewsGroup dataset with a popular classifier learning method, support vector machine (SVM). Our weighting scheme ite achieved the best result in classification accuracy compared with other existing methods. And ite has the most stable performance with the reduction of training samples as well. Furthermore, our method has a built-in property to prevent over-weighting in STW. Over-weighting is a newly proposed concept especially with supervised term weightings in our earlier work and re-introduced here. Caused by the improper singular terms and too large ratios between term weights, over-weighting could deprive the performance of text classification tasks.

In this paper, we propose a sentence encoding-based model for recognizing text entailment (RTE). In our approach, the encoding process of sentences consists of two stages. Firstly, average pooling is used over word-level bidirectional LSTM (biLSTM) to generate a first-stage sentence representation. Secondly, attention mechanism is employed to replace average pooling on the same sentence for better representations. Instead of using target sentence to attend words in source sentence, we utilize the sentence’s first-stage representation to attend words appeared in itself, which is called “Inner-Attention” in our paper. Experiments conducted on Stanford Natural Language Inference (SNLI) Corpus has proved the effectiveness of “Inner-Attention” mechanism. With less number of parameters, our model outperformed the existing best sentence encoding-based approach by a large margin.

This paper designs and implements a smart garden system that can be used inside the house. To do this, we used particle filter and environmental sensors, which are open hardware controllers, and designed to control and observe automatic water supply, lighting, and growth monitoring with three wireless systems (Bluetooth, Ethernet, Wi-Fi). This system has been developed to make it possible to use it in an indoor space such as an apartment, rather than a large-scale cultivation system such as a conventional plant factory which has already been widely used. The developed system collects environmental data by using soil sensor, illuminance sensor, humidity sensor and temperature sensor as well as control through smartphone app, analyzes the collected data, and controls water pump, LED lamp, air ventilation fan and so on. As a wireless remote control method, we implemented Bluetooth, Ethernet and Wi-Fi. Finally, it is designed for users to enable remote control and monitoring when the user is not in the house.

Using typical algorithms for training multilayer perceptron (MLP) creates some difficulties like slow convergence speed and local minima trapping in the solution space. Bio-inspired learning algorithms are famous for solving linear and nonlinear combinatorial problems. Artificial Bee Colony (ABC) algorithm is one among the famous bio-inspired algorithms. However, due to slow exploration process, it has been focused by researchers for further enhancement in optimization area. Therefore, this paper proposed a new hybrid swarm based learning algorithm called Global Crossover Artificial Bee Colony (GCABC) algorithm for training MLP for solving boolean classification problems. The simulation results of proposed GCABC algorithm compared with standard bio-inspired algorithms such as ABC, and Global Artificial Bee Colony (GABC) show that the proposed algorithm is achievable and efficient results in benchmark boolean function classification, with fast convergence speed.

After analyzing the disadvantages of the classical K-means clustering problem, an improved cuckoo search algorithm (ICS) is applied to cluster analysis, and this paper proposes a novel hybrid clustering algorithm based on Genetic algorithm (GA). The hybrid algorithm includes two modules. At the initial stage, the cuckoo search algorithm (CS) is executed, the clusters’ result are used to the crossover and mutation of genetic algorithm for local search. Comparision of the performance of the proposed approach with the cluster method based on CS and GA algorithm are experimented. The experimental result show the proposed meth has not only higher accuracy bust also higher level of stability. And the faster convergence speed can also be validated by statistical results.

The 0–1 knapsack problem (0–1KP) is a well-known combinatorial optimization problem. It is an NP-hard problem which plays significant roles in many real life applications. Dragonfly algorithm (DA) a novel swarm intelligence optimization algorithm, inspired by the nature of static and dynamic swarming behaviors of dragonflies. DA has demonstrated excellent performance in solving multimodal continuous problems and engineering optimization problems. This paper proposes a binary version of dragonfly algorithm (BDA) to solve 0–1 knapsack problem. Experimental results have proven the superior performance of BDA compared with other algorithms in literature.

Moth Swarm Algorithm (MSA) is a new swarm intelligent algorithm, it is inspired by the moth looking for food, phototaxis and celestial navigation in the dark environment, proposed a moth search algorithm. Because the algorithm has good convergence speed and high convergence precision, it is applied in many fields. Cluster analysis, as an effective tool in data mining, has attracted widespread attention and has been developed rapidly and has been successfully applied in recent years. Among the many clustering algorithms, the K-means clustering algorithm is easy to implement, so it is widely used. However, the K-means algorithm also has the disadvantages of large computational complexity and clustering effect depending on the selection of the initial clustering center, which seriously affects the clustering effect, and the algorithm is easy to fall into the local optimum. To solve these problems, The MSA is applied to cluster analysis, the results show that the MSA not only achieves superior accuracy, but also exhibits a higher level of stability.

The recently proposed twin parametric insensitive support vector regression, denoted by TPISVR, which solves two dual quadratic programming problems (QPPs). However, TPISVR has at least four regularization parameters that need regulating. In this paper, we increase the efficiency of TPISVR from two aspects. Fist, we propose a novel least squares twin parametric insensitive support vector regression, called LSTPISVR for short. Compared with the traditional solution method, LSTPISVR can improve the training speed without loss of generalization. Second, a discrete binary particle swarm optimization (BPSO) algorithm is introduced to do the parameter selection. Computational results on several synthetic as well as benchmark datasets confirm the great improvements on the training process of our LSTPISVR.

The BoF (bag-of-features) model is one of the most famous models applied to many fields in computer vision and has achieved impressive results. However, the SIFT/HOG visual words have a limit discriminative power which is partly due to the fact that it only describes the local gradient distribution. In the meanwhile, there is still redundancy and hidden information existed in the formed histogram. Considering these respects, we propose a multi-modal SPM model which fuses global features to complement traditional local ones and conducts dimensionality reduction in local spaces for mining possible feature dependencies. Experimental results show the efficiency of the proposed method in comparison with the existing counterparts.

Traditional image steganography modifies the content of the image more or less, it is hard to resist the detection of image steganalysis tools. New kind of steganography methods, coverless steganography methods, attract research attention recently due to its virtue of do not modify the content of the stego image at all. In this paper, we propose a new coverless steganography method based on robust image hashing. Firstly, we design an effective and stable image hash by using the orientation information of the SIFT feature points. Then the local image database is created and the corresponding hash values of these images in the database are computed. Secondly, the secret message is divided into segments with the same length as the hash sequences. And a series of images are chosen from the image database by matching the secret information segments and the hash sequences of all the images. Finally, these images are transmitted as the carriers of the secret information. When the receiver receives these images, the secret information is extracted by using the shared hash method. Due to the characteristics that SIFT features can resist common image attacks in a certain extent, the secret information corresponding to the hash has strong robustness. To improve the retrieval and matching efficiency of the hashing system, an inverted index of quadtree structure is designed. Compared with the traditional image steganography, this method does not modify the content of the image itself, therefore, can effectively resist steganalysis tools. Furthermore, we compare the proposed method with state-of-art coverless steganography method which also based on image hash, and experimental results show that our method has higher capacity, robustness and security than the method proposed in [15].

As the commonest part of social networks, sharing an image in social networks transmits not only can provide more information, but also more intuitive than any text. However, images also can leak out information more easily than text, so the audit of image content is particularly essential. The disclosure of a tiny image, which involves sensitive information about individual, society even the state, may trigger a series of serious problems. In this paper, we design an image firewall to detect sensitive image content through joint sparse representation on features. We take LBP, SIFT and Wavelet features into consideration, trying to find an effective combination among these features. We also find some features, which have the same accuracy but less time cost. In addition, we consider the spatial relation of the detected objects, especially the distance between the persons appeared in an image. Experimental results show the effectiveness of the proposed methods.

To reconstruct the appearance of an unknown skull based on knowledge, the most similar skull should be retrieved from the skull Database. The process is called skull registration. A hierarchical skull registration method with bounded rotation angle is proposed in this paper. The surface of the skull is divided into concave or convex regions by K-means. The optimal 3D transform is searched for each potential pair of matched feature regions approximately to align the skulls roughly. And then the novel ICP (Iterative Closest Point) algorithm with a bounded rotation angle (BRA-ICP) is applied for fine registration. To show the generality, the proposed algorithm is applied not only for skull registration but also for the public data registration. Experiments show that the proposed algorithm can achieve better registration accuracy and higher iterative convergence speed in the fine registration stage and the entire process is completed without human intervention.

Sex determination is the first step in criminal investigation. Due to its easiness for protection, skull is considered to be the second important skeleton for sex determination. However, not all criminal cases can provide complete skull. In this paper, we present a sex determination model for incomplete skull. First, the skull is divided into seven partitions, the feature points are marked and the unmeasurable features are quantized. Then, the optimal feature subset of each partition is selected by using forward stepwise regression method based on maximum likelihood estimation. Seven partition sex determination decision models were set up and tested by using leave-one-out test. Finally, the final sex determination for incomplete female and male skull were constructed. Experiments show that any 3 partitions are enough to determine the sex of a skull with a high accuracy.

We propose a super-pixel segmentation algorithm based on normalized Euclidean distance for handling the uncertainty and complexity in medical image. Benefited from the statistic characteristics, compactness within super-pixels is described by normalized Euclidean distance. Our algorithm banishes the balance factor of the Simple Linear Iterative Clustering framework. In this way, our algorithm properly responses to the lesion tissues, such as tiny lung nodules, which have a little difference in luminance with their neighbors. The effectiveness of proposed algorithm is verified in The Cancer Imaging Archive (TCIA) database. Compared with Simple Linear Iterative Clustering (SLIC) and Linear Spectral Clustering (LSC), the experiment results show that, the proposed algorithm achieves competitive performance over super-pixel segmentation in the state of art.

Keratoconus is a bilateral progressive corneal disease characterized by thinning and apical protrusion; its early diagnosis is fundamental since it allows one to treat this rare disease by cross-linking approach, thus preventing a major corneal deformation and avoiding more invasive and risky surgical therapies, such as cornea transplant. Ophthalmology improvements have allowed a more rapid, precise and painless acquisition of corneal biometric parameters which are useful to evaluate alterations and abnormalities of eye’s outer structure. This paper presents a study about Keratoconus diagnosis based on a machine learning approach using corneal physical and morphological parameters obtained through Precisio™ tomographic examination. Artificial Neural Networks (ANNs) have been used for classification; in particular, a mono-objective Genetic Algorithm has been used to obtain the best topology for the neural classifiers for different input datasets obtained from features ranking. High levels of accuracy (higher than 90%) have been reached for all types of classification; in particular, binary classification has showed the best discrimination capability for Keratoconus identification.

This paper presents a reversible data hiding algorithm without intra-frame distortion drift based HEVC. We embed the secret information into the multivariate array of the 4 × 4 luminance DST blocks to avert the distortion drift. With the inverse operation of multivariate array in decoder, the embedded video is perfectly reconstructed as the original encoded video. Moreover, the entire process of embedding and extracting is very easy to operate. The superiority of the presented algorithm is verified through experiments.

A little link and wavelength failure of the optical network would result in a great traffic loss. Therefore a well-performed protection strategy is of great importance. We present a wavelength protection algorithm for the optical network. Compared with the previous works, this method achieves better protection effect for the wavelength than the algorithms proposed ever before, and this algorithm provides a higher wavelength utilization ratio and lower wavelength blocking probability.

Computer music humming tone modulation technology is mainly to enhance the computer music sound of human nature. Such as enhancing the authenticity of the sound and beauty or enhance the performance of the human and virtual sound field effects, etc. It mainly uses the existing technical means of hardware and software to change the basic attributes of sound, which mainly includes the sound envelope and the virtual sound field effect modulation. Through the modulation of the preset tone, you can improve the characteristics of human voice performance.

This paper presents a readable H.265/HEVC data hiding algorithm. To avert intra-frame distortion drift, we first give the Condition of the directions of intra-frame prediction. Then we embed the message into the multi- coefficients of the 4 × 4 luminance DCT blocks of the selected frames which meet the Condition. The experimental results show that this data hiding algorithm can effectively avert intra-frame distortion drift, and get good visual quality.

Evaluation model based on attribute coordinate has made some achievements in both theoretical research and practical applications. However, if the new evaluation samples are added, the evaluation model needs to be reconstructed rather than the dynamic updating. Almost no progress has been made on how to dynamically update the evaluation model. Thus, this paper puts forward a dynamic updating algorithm based on barycentric coordinates and satisfaction function to effectively solve this problem. The experiment results show the reasonability and effectiveness of this algorithm.

One of the current challenges in machine learning is to develop intelligent systems that are able to learn consecutive tasks, and to transfer knowledge from previously learnt basis to learn new tasks. Such capability is termed as lifelong learning and, as we believe, it matches very well to counter current problems in cybersecurity domain, where each new cyber attack can be considered as a new task. One of the main motivations for our research is the fact that many cybersecurity solutions adapting machine learning are concerned as STL (Single Task Learning problem), which in our opinion is not the optimal approach (particularly in the area of malware detection) to solve the classification problem. Therefore, in this paper we present the concept applying the lifelong learning approach to cybersecurity (attack detection).

Lie speech detection is a typical psychological calculation problem. As the lie information is hidden in speech flow and cannot be easily found, so lie speech is a complex research object. Lie speech detection is not only need to pay attention to the surface information such as words, symbols and sentence, it is more important to pay attention to the internal essence structure characteristics. Therefore, based on the study of speech signal sparse representation, this paper proposes a Stack Sparse Automatic Encoder (SSAE) deep learning model for lying speech characteristics extraction. The proposed method is an effective one, it can reflect people’s deep lying characteristics, and weaken lying person’s personality traits. The deep characteristics compensate the lack of lie expression of basic acoustic features. This improved the lying state correct recognition rate. The experimental results show that, due to the introduction of deep learning characteristics, the individual lying recognition rate has increased by 4%–10%. This result suggests that, the lie detection based on speech analysis method is feasible. Furthermore, the proposed lying state detection based on speech characteristic provides a new research way of psychological calculation.

In this paper, we propose a novel approach to detect fire based on convolutional neural networks (CNN) and support vector machine (SVM) using tensorflow. First of all, we construct a large number of different kinds of fire and non-fire images as the positive and negative sample set. Next we apply Haar feature and AdaBoost cascade classifier to extract the region of interest (ROI). Then, we use CNN-SVM to filter the results of Haar detection and reduce the number of negative ROI. The CNN is constructed to train the dataset with four convolutional layers. Finally, we utilize SVM to replace the fully connected layer and softmax to classify the sample set based on the training model in order. Experimental results show that the method we proposed is better than other methods of fire detection such as CNN or SVM etc.

The stochastic gradient descend (SGD) is a prevalence algorithm used to optimize Convolutional Neural Network (CNN) by many researchers. However, it has several disadvantages such as occurring in local optimum and vanishing gradient problems that need to be overcome or optimized. In this paper, we propose a hybrid learning algorithm which aims to tackle the above mentioned drawbacks by integrating the methods of particle swarm optimization (PSO) and SGD. To take advantage of the excellent global search capability of PSO, we introduce the velocity update formula which is combined with the gradient descend to overcome the shortcomings. In addition, due to the cooperation of the particles, the proposed algorithm helps the convolutional neural network dampen overfitting and obtain better results. The German traffic sign recognition (GTSRB) benchmark is employed as dataset to evaluate the performance and experimental results demonstrate that proposed method outperforms the standard SGD and conjugate gradient (CG) based approaches.

Improving average effectiveness is an objective of paramount importance of ranking model for the learning to rank task. Another equally important objective is the robustness—a ranking model should minimize the variance of effectiveness across all queries when the ranking model is disturbed. However, most of the existing learning to rank methods are optimizing the average effectiveness over all the queries, and leaving robustness unnoticed. An ideal ranking model is expected to balance the trade-off between effectiveness and robustness by achieving high average effectiveness and low variance of effectiveness. This paper investigates the effectiveness-robustness trade-off in learning to rank from a novel perspective, i.e., the bias-variance trade-off, and presents a unified objective function which captures the trade-off between these two competing measures for jointly optimizing the effectiveness and robustness of ranking model. We modify the gradient based on the unified objective function using LambdaMART which is a state-of-the-art learning to rank algorithm, and demonstrate the strategy of jointly optimizing the combination of bias and variance in a principled learning objective. Experimental results demonstrate that the gradient-modified LambdaMART improves the robustness and normalized effectiveness of ranking model by combining bias and variance.

The basic idea of ssHC is to leverage domain knowledge in the form of triple-wise constraints to group data into clusters. In this paper, we perform extensive experiments in order to evaluate the effects of different distance metrics, linkages measures and constraints on the performance of two ssHC algorithms: IPoptim and UltraTran. The algorithms are implemented with varying proportions of constraints in the different datasets, ranging from 10% to 60%. We found that both IPoptim and UltraTran performed almost equally across the seven datasets. An interesting observation is that an increase in constraint does not always show an improvement in ssHC performance. It can also be observed that the inclusion of too many classes degrades the performance of clustering. The experimental results show that the ssHC with Canberra distance perform well, apart from ssHC with well-known distances such as Euclidean and Standard Euclidean distances. Together with complete linkages and small amount of constraints of 10%, ssHC can achieve good results of an F-score close to 0.8 and above for four out of the seven datasets. Moreover, the output of non-parametric statistical test shows that using the UltraTran algorithm in combination with the Manhattan distance metric and Ward.D linkage method provides the best results. Furthermore, utilizing IPoptim and UltraTran with the Canberra distance measure performs better for the given datasets.

Rotation forest (RoF) is an ensemble classifier based on the combination of linear analysis theories and decision tree algorithms. In existing works, the RoF has demonstrated high classification accuracy and good performance with a reasonable number of base classifiers. However, the classification accuracy drops drastically for linearly inseparable datasets. This paper presents a hybrid algorithm integrating kernel principal component analysis and RoF algorithm (KPCA-RoF) to solve the classification problem in linearly inseparable cases. We choose the radial basis function (RBF) kernel for the PCA algorithm to establish the nonlinear mapping and segmentation for gene data. Moreover, we focus on the determination of suitable parameters in the kernel functions for better performance. Experimental results show that our algorithm solves linearly inseparable problem and improves the classification accuracy.

As an important public transportation, Taxi is used for passengers every day, which is one of the primary causes for traffic jams. For passengers, knowing the difficulty degree of taking a taxi at a particular time and place can help us plan the journey effectively. Nevertheless, the existing predication models for traffic are not able to express the difficulty degree of choosing a taxi. In order to solve this problem, we can use historical data of taxi status to analysis and predict the possibility of taxi-hailing at a specific time and place. In this paper, we present a classification and predication framework for taxi-hailing. In this framework, firstly we use K-Means clustering algorithm to divide the taxi data into different clusters. Then we use Echarts to extract the features of each cluster in order to show the different difficulty degree. Next we use neural network to generate the predication result using the result of K-Means. On this basis, we propose a method to make the predication of taxi-hailing at a particular time and place, which can calculate the possibility score of taxi-hailing. Finally, we make a prediction using this framework and compare the predication results with the actual travelling data report. The comparison results verify the reliability of this framework.

In this paper, we propose a real-time recommender system (PRACE) for taxi drivers to find a next passenger and start a new trip efficiently, based on historical GPS trajectories of taxis. To provide high-quality passenger-seeking advice, PRACE takes passenger prediction, road condition estimation, and earnings into ranking simultaneously. Different from many previous researchers, we not only pay more attention to the driving context of taxis (i.e., driving directions, positions, etc.) but also extract meaningful representations of these attributes, using deep neural networks. To enhance the effect of learning, the result of statistics is added to the input of models. Relying on the map meshing method, we treat the prediction task as a multi-classification problem rather than a regression problem and make comparisons with several state-of-the-art methods. Finally, we evaluate our method through extensive experiments, using GPS trajectories generated by more than 10,000 taxis from the same company over a period of two months. The results verify the effectiveness, efficiency, and availability of our recommender system.

The problem of matrix approximation appears ubiquitously in recommendation systems, computer vision and text mining. The prevailing assumption is that the partially observed matrix has a low-rank or can be well approximated by a low-rank matrix. However, this assumption is strictly that the partially observed matrix is globally low rank. In this paper, we propose a local sensitive formulation of matrix approximation which relaxes the global low-rank assumption, leading to a representation of the observed matrix as a weighted sum of low-rank matrices. We solve the problem by nonconvex optimization which exhibits superior performance of low rank matrix estimation when compared with convex relaxation. Our experiments show improvements in prediction accuracy over classical approaches for recommendation tasks.