ABSTRACT. Network Functions Virtualization (NFV) is a recent trend of network transformation that helps service providers offer new and multiple services in a more agile and cost effective way. However, the softwarization and cloudization of network functions can result in high congestion and low network performance for virtual networks. We develop a multipath routing solution for minimizing the maximum link utilization, thus leading to the performance improvement of NFV-based systems, as well as the efficient utilization of network resources. Our proposed algorithm can adapt the link weight system to the dynamic change of service demands for optimizing the distribution of network traffic, with regard to the fundamental NFV characteristics and the Equal-cost Multipath (ECMP) routing feature. The evaluation results show that the proposed solution outperforms a multipath solution with fixed link weight on various performance metrics. Interestingly, we find that the network performance is not significantly improved by increasing the number of paths beyond a threshold.

ABSTRACT. High performance computing (HPC) clouds provide services that users can run their applications on cloud infrastructure. These systems consume a lot of energy (kWh); therefore reducing energy consumption is a high priority for any cloud provider. This paper studies application of vector bin packing heuristic and Neural Network (NN) to allocate virtual machines (VMs) onto physical machines (PMs) that minimizes total energy consumption of the physical machines. In our scenario, a list of virtual machines from request queue needs to assign to system for every interval time (T) and minimize total energy consumption. Based on Best Fit Decreasing heuristic, we proposed Best Fit Decreasing Neural Network (BFD-NN), which contains an evaluation function (f) that finds the most efficient physical machine for each VM from requests in system. We also proposed a process to optimize weight of coefficients in f for every request which users submit to the system based on information of users’ requests in the past by using Parallel Genetic Algorithm (PGA) and Neural Network. Our method is a new approach because it not only uses knowledge from requests of users but also considers time dimension of virtual machines. Two job parallel workload models in Parallel Workloads Archive are used to evaluate our approach. The simulation results illustrate that BFD-NN could reduce up to 15% total energy consumption compared with state-of-the-art heuristics (such as Best Fit and First Fit Decreasing) in online allocation virtual machines.

ABSTRACT. Cloud computing is a virtualized compute power and storage delivered via platform-agnostic infrastructures of abstracted hardware and software accessed over the Internet. Workflow is adopted as an attractive paradigm for its powerful ability in expressing a wide range of applications, including scientific computing, multi-tier Web, and big data processing applications. Despite it has been the focus of many researchers, a handful efficient solutions have been proposed for Cloud computing. In this work, we propose a novel algorithm for workflow scheduling that is derived from the Opposition-based Differential Evolution method, MODE. This algorithm not only ensures fast convergence but also averts getting trapped in local extrema. Our simulation experiments CloudSim show that MODE is superior to its predecessors. Moreover, the deviation of its solution from the optimal one is negligible.

ABSTRACT. Legacy distributed applications are moving to the cloud in order to reduce operational costs and provide more scalable services to users. These applications are built by inter-connecting independently developed packages and services distributed over multiple machines. The installation, deployment and management of such applications on the cloud is hard, and usually performed either manually or by writing customized scripts. This is highly error-prone and it has been shown that misconfiguration is the root of many errors. We present AutoBot, a system for configuring, installing, and managing complex legacy application stacks deployed on the cloud which dynamically evolve in time. AutoBot consists of three components: a novel and simple model describing components’ configuration properties and inter-dependencies; a dynamic deployment and configuration protocol ensuring the correct resolution of these inter-dependencies; a runtime system that guarantees the correct deployment of the application across multiple machines as well as the management of the deployed system. This reduces the tedious and error-prone manual process of application configuration, installation, and management. We have implemented and used AutoBot to successfully host a number of applications. In this paper, we specifically describe our experiences in using AutoBot to manage a traditional N-tier Java application in the cloud. In order to check that this protocol works as expected, we also formally specify and verify some interesting properties.

ABSTRACT. Identifying the interactions between drugs and targets is a crucial step in the process of discovering new drugs. There has been a number of computational methods proposed for the problem. Among them, machine learning-based methods usually utilizes the similarity between drugs and between targets to build kernel matrices, which are used to predict novel drug-target interactions with classification models. While network-based methods usually formulate the prediction as a ranking problem where candidate targets are according to a drug of interest and/or its known targets. A common disadvantage of the network-based methods is that they mainly look for novel targets which are close to known targets in the network. In this study, we proposed a method, namely SigTarget, to overcome this limitation. More specifically, SigTarget ranks candidate targets based on a probability with which they connect to known targets by choosing significant links between known and candidate targets. This method was adapted from an algorithm calculating relative importance between nodes in a network. Simulation results show that SigTarget was better than some existing methods such as TBSI, DBSI and RWR for a set of drugs collected from KEGG database. In addition, we showed the ability of SigTarget in predicting novel drug targets by showing that highly ranked candidate targets obtained from SigTarget are also verified in another drug database, DrugBank.

ABSTRACT. Today, more and more Internet users are willing to share their feeling, activities, and even their intention about what they plan to do on online social media. We can easily see posts like "I plan to buy an apartment this year", or "We are looking for a tour for 3 people to Nha Trang" on online forms or social networks. Recognizing those user intents on online social media is really useful for targeted advertising. However fully understanding user intents is a complicated and challenging process which includes three major stages: user intent filtering, intent domain identification, and intent parsing and extraction. In this paper, we propose the use of machine learning to classify intent-holding posts into one of several categories/domains. The proposed method has been evaluated on a medium-sized collections of posts in Vietnamese, and the empirical evaluation has shown promising results with an average accuracy of 88%.

ABSTRACT. Credit scoring is a fundamental and one of the most complex tasks that financial institutions have to deal with. Commonly this problem is considered as default probability prediction for the lenders counterparts. Knowing predictors that significantly contribute to default prediction has recently emerged as a crucial issue of credit risk analysis. Default prediction and default predictor selection are two related issues, but many existing approaches address them separately. In this paper a unified procedure is proposed that based on the regularization approach with logistic regression as an underlying model, which simultaneously selects the default predictors and optimizes all the parameters within the model. We give the strong probabilistic statement of shrinkage criterion for features selection. The proposed regularization do not corrupt the relevant predictors, can select correlated predictors into model, gives stable subset of relevant features. Experimental results show that the proposed framework is competitive on both artificial data and publicly available data sets.

ABSTRACT. In any software development life cycle, requirement and software changes are inevitable. One of the factors that influences the effectiveness of the change acceptance decision is the accuracy of the effort prediction for requirement changes. There are two current models that have been widely used to predict rework effort for requirement changes which are algorithmic and non-algorithmic models. The algorithmic model is known for its formal and structural way of prediction and best suited for Traditional software development methodology. While non-algorithmic model is widely adopted for Agile software development methodology of software projects due to its easiness and requires less work in term of effort predictability. Nevertheless, none of the existing effort prediction models for requirement changes are proven to suit both, Traditional and Agile software development methodology. Thus, this paper proposes an algorithmic-based effort prediction model for requirement changes that uses change impact analysis method which is applicable for both Traditional and Agile software development methodologies. The proposed model uses a current selected change impact analysis method for software development phase. The proposed model is evaluated through an extensive experimental validation using case study of six real Traditional and Agile methodologies software projects. The evaluation results confirmed a significance accuracy improvement of the proposed model over the existing approaches for both Traditional and Agile methodologies.

ABSTRACT. This paper presents a fully automated and flexible ConvNet-based classifier for multi-label image annotation. The classifier alleviates hierarchical representation of image from a convolutional neural network, and adaptive thresholding technique on the ranked list of label scores. The method can annotate images with arbitrary number of labels that the classifier finds fit, as opposed to common methods which only assign a fixed number. Results show state-of-the-art on classification accuracy and competitive annotation performance across 2 intrinsically different datasets, Corel5K and MSRCv2. Although the proposed method shows some limitation in learning label semantics, empirical study indicates that it was due to the established drawback of univariate loss function, which the classifier optimised, in multi-label classification. It, therefore, opens for number of directions to improve the performance, and still retains the merits of the proposed method.

ABSTRACT. Dictionary learning (DL) has been successfully applied to various pattern recognition problems. Some supervised dictionary learning method integrate a classifier training with the dictionary learning model to make better accuracy outcome. However, it is the computational complexity of the sparsity constrain make learning problem expensive. There are some proposed methods come out to relax the problem using $l_2$-norm. The new approaches have shown theirs advantages in both accuracy and classification time. In this paper, we present an combination of studying a pair dictionary: synthesis dictionary for representation task and analysis dictionary for classification task with a classifier training term in order to make the classification process just only depend on the analysis dictionary and the classifier. The proposed method has a high accuracy but the classification time is extremely fast. We experimented on three datasets of image classification and the model show promising results compare with other state-of-the-art methods.

ABSTRACT. In this paper, we propose a new solution to identify hand gestures of Vietnamese Sign Language (VSL) via a sequence of images (video) collected from the depth sensor in a Microsoft Kinect. First, a preprocessing is performed to localize and separate the hand from each image and then remove possible noise. In the next stage, the object is extracted to select key frames, which support to represent a segment of the video. Each determined key frame is then converted to a binary image and estimate some biological information such as the hand boundary, finger positions and the palm center. The position of palm centre and fingertips are also localized in 3D space. The process of recognition is performed using Support Vector Machine (SVM) method. The experiments show that the proposed approach is promising since the recognition accuracy is about 91%.

ABSTRACT. A combination of Conformal Geometric Algebra and Principal Components Regression methods showed the effect of recognition for 3D rotation objects. The advantage of this method is using of the distribution hyper-spheres in order to classify the 3D rotation objects. This study focuses on applying for Human Activity Recognition. However, using the 3D coordinates of the markers for mounting experimentally on the human body can meet a difficulty to generate the distribution of hyper-spheres although this marker may orbit around the other ones. Therefore, feature selection to data of a class to be distributed hyper-spheres is necessary. This study calculated the correlation between markers of objects, from which to build a set of important markers to extract feature to increase the accuracy of classification rate. Through experiments, it shows that the proposed method of marker selection provides the better results in classification than using all markers.

ABSTRACT. Emotions, these reflexes that push human beings to make decisions quickly and without a deep and clear reasoning process, have been considered for a long time contrary to any rational reasoning process. Only recently the key role of emotions in decision-making process has been highlighted. In this article, we focus on fear related emotions and their positive impact on the survival capabilities of human beings in case of crisis situation. The purpose of this paper is to clarify the impact of emotion on agents in the evacuation process; we only focus here on emotions and not on the evacuation process itself. We formalize the influence of three factors (decay, environment, contagion) on agents emotion and explore the influence of each of them. The entire theoretical model has been implemented in the simulation platform GAMA.

ABSTRACT. In this work, we present a paradigm to build a decision support tool which enable users to "play" with the model, change the input parameters, test different intervention scenarios,... in case of infectious diseases. The proposed paradigm involves modeling of closed to realistic social contact networks through which the disease transmissions take place as well as the social influence networks which take part in determining individual decisions and hence, lead to impacts on the propagation of the disease. The proposed model follows agent based approach and uses statecharts for modeling of agent behaviors focusing on the behaviors of each individual when he/she is being sick or when hearing of the sickness of the others. The dynamics of attitude of a person (which affects that agent's decision) is characterized using opinion dynamics model that take place on those influence networks. Different interventions strategies are also investigated that take into account the fact that agent is incentive. Our preliminary simulation results shows that, with the making best use of available (demographic, transportation,...) data and considering of behaviors at individual level, the proposed model can capture the spreading characteristics of the disease as well as the impacts of intervention strategies on its propagation process at the whole population level.

ABSTRACT. Public transportation modes like buses play an important role in our real life. Designing a good bus system can bring advantages to passengers and the society. One of key design criteria is the connectivity which can be measured by the minimum number of bus lines a passenger must take to travel from one station to another. This paper considers the bus system design problem with the objective of improving the connectivity of the system from the existing infrastructure. We define the problem in the form of an optimization problem and propose an algorithm for the solution. Results obtained from the real instances derived from the bus system of Ho Chi Minh city are analysed and reported.

ABSTRACT. In many big cities, the volume of daily municipal solid waste generation is huge. The management of solid waste collection and transportation is a challenge. Actually, in many developing countries, this work mainly depends on experiences. Therefore, it is not effective; the waste retention time at collection points is long. The purpose of our work is to explore a decision support system for municipal waste collection and transportation by integrating GIS map, smart devices and agent-based model. Based on the real-time data that is collected and transmitted by smart devices, the decision support system calculates the optimized path for each vehicle. A case study related to Hagiang city, Vietnam, is presented to show the efficiency of the proposed model. From the optimized results, it has been found that the cost of the MSW collection is reduced.

ABSTRACT. Video instance search or also well known as object retrieval is a fundamental task in computer vision field and has a lot of applications. Most state-of-the-art systems are based on the Bag-of-Words model (BOW) for representing video frames and target object. When searching on nearly planar and rich-textured objects such as buildings and book cover, BOW argue to be a suitable model with very high performance. However, when searching on harder but more popular objects such as a specific person, BOW model still keep a lower performance. In this paper, we consider a new type of query which covers wo most popular topics: searching a person in a specific location. Inspired by recent successes of deep learning techniques, we propose new framework which leverage the powerful of both BOW model and deep feature in instance search. In particular, we use a linear kernel classifier instead of using L2 distance to compute similarity between two deep features. For further improvement, scene tracking are employed to deal with the cases face of query person is not detected. To evaluate the proposed methods, we conduct experiments over a standard benchmark dataset (TRECVID Instance Search 2016) with more than 300 GB in storage and 464 hours in duration. The results show that, our proposed methods significant improve the baseline system.

ABSTRACT. Person tracking is crucial in any automatic person surveillance systems. In this problem, person localization and re-identification (Re- ID) are both simultaneously processed to show separated trajectories for each individual. In this paper, we propose to use mixture of WiFi and camera systems for person tracking in indoor surveillance regions covered by WiFi signals and disjointed camera FOVs (Field of View). A fusion method is proposed to combine the position observations achieved from each single system of WiFi or camera. The combination is done based on an optimal assignment between the position observations and predicted states from camera and WiFi systems. The correction step of Kalman filter is then applied for each tracker to give out state estimations of locations. The fusion method allows tracking by identification in non-overlapping cameras, with clear identity information taken from WiFi adapter. The experiments on a multi-model dataset show outperforming tracking results of the proposed fusion method in comparison with vision-based only method.

ABSTRACT. Falling when being alone is the most dangerous circumstance with the elderly. To resolve the safety for older people living alone, this paper proposes a method of indoor fall detection using single camera system. Falls are detected based on the analysis of motion orientation, motion magnitude, and human shape changes. Our method has been experimented on Li2e Fall detection Datasets, including 4 sections corresponding to 4 different environments: Home, Lecture room, Office and Coffee room. The datasets have a total of 221 videos presenting 126 falls with many different scripts, and the remaining representing daily activities such as moving, sitting, lying and floor cleaning, etc. The experimental results exhibit the high detection accuracies and very fast processing capability.

ABSTRACT. In this paper, we propose an algorithm for hiding DNA sequence in H.264/AVC video. The DNA sequence is encrypted by a rule before hiding, then we embed encrypted DNA sequence into the quantized discrete cosine transform (QDCT) coefficients of the 4×4 macroblocks (MBs) of Intra-frames (I-frames). To avert the distortion drift, we use the directions of I-frame prediction. By using all paired-coefficients that meet -1, 0 and 1, we do not need use any threshold for limiting the capacity of embedded data. The experimental results show that the proposed algorithm has improved with other existing algorithms in term embedding capacity.

ABSTRACT. A Lens Free Imager (LFI) is one of promising candidates for massive visual inspection systems in biology and medical science. In a practical application, millions of samples need to be inspected in a short period of time, which a single Lens Free Imager (LFI) system cannot accomplish. Therefore, we propose a high performance parallel LFI system called RALFIE (Real-time Advanced LFI Evaluation system). In this paper, we introduce the concept and detail of the design of RALFIE and a target application.

ABSTRACT. The emergence of advanced network technologies and decreasing DRAM costs have enabled data centers to deploy key-value in-memory stores to deliver ultra-low latency data service for real-time applications. Although key-value model helps in-memory stores to simplify their query processing and data management mechanisms, its simplicity have applications to decompose generate multiple item lookups to perform simple requests such as constructing a HTML files. Those requests are often executed sequentially and seriously affect the overall performance. In this paper, we introduce KELI, a novel in-memory key-value store that employs Key-value-with-Links, a simple but quite efficient data model, to solve this issue. On the top of key-value model, KELI establishes links between items so instead of invoking expensive item lookups, it could simply follows those links to retrieve needed data. KELI also uses RDMA Read to implement link chasing to minimize searching overhead. Some experiments have been conducted to examine the performance of KELI. Their results indicate that KELI runs three times faster than Redis for data lookups. Furthermore, KELI runs seven times faster than Redis when handling real workloads generated by a practical application.

ABSTRACT. Mobile phone is being well integrated into people's daily life. Due to a large amount of time spending with them, users expect to have a good experience for their daily tasks. The mobile operating system's scheduler is in charge of distributing CPU computational power among these tasks. However, it currently has not yet taken into account dynamic frequencies of CPU cores at runtime. This unawareness of the scheduler with CPU frequency increases unresponsiveness of user interface to user interactions, and consequently reduces user experience on using mobile devices. In this paper, we propose an extension of process scheduler which takes into account the dynamic CPU frequency when scheduling the tasks. Our method increases smoothness of user interface to user interactions by lowering and stabilizing interface frame times. Experimental results show that our proposed scheduler reduces amount of frame time peaks up to 40%, which helps greatly in improving user experience on mobile devices.

ABSTRACT. Research in mobile ad-hoc networks (MANETs) has been in-progress towards real-time implementation, such as routing protocols, real wireless ad-hoc routers, as well as device drivers for wireless network cards. However, little work focuses on the real-time architecture and open-source code analysis of components in MANETs node, namely the wireless ad-hoc router (WAHR). This paper, in contrast, concentrates on such an issue, i.e., the open-source framework for integration of open-source wireless ad-hoc drivers into the Linux kernel. In particular, we presents and discusses deeply the operation and the integration of Atheros ath5k and Broadcom b43 device drivers into the Linux kernel to implement a soft, real WAHR for the deployment of real-time ad-hoc applications. The open-source code of Atheros ath5k and Broadcom b43 device drivers are analyzed and discussed on different flows, including: i) Configuration Path; ii) Reception (RX) Path; and iii) Transmission (TX) Path. Finally, realistic scenarios are used to verify for the operations of ath5k/b43 device drivers and WAHR

ABSTRACT. In this work, we present procedures for image denoising based on dynamic programming procedure for maximum a posteriori probability estimation. A new non-convex type regularization is used, with ability to flexibly set a priori preferences, using different penalties for various ranges of differences between the values of adjacent image elements. Proposed procedures can take into account heterogeneities and discontinuities in the source data.

ABSTRACT. In this paper, we propose a method to remove noise in RGB-color images. This method is based on a total variation of intensity function of images. We develop our method proposed before to remove a linear combination of Gaussian and Poisson noises in grayscale images. Our denoising model “feels” well the wide range of proportion of two types of noises and appears to be the good basis for the real case to remove the Poisson-Gaussian noise superposition. This combined noise can be used to well approximate real noises in color raster images.

ABSTRACT. Considering that the existing methods of cell segmentation are sensitive to noise, in this paper, a novel method was proposed to obtain the contour of cervical cell accurately and robustly. Firstly, the meanshift algorithm was utilized to smooth the cell image. Then, the initial contour of cervical cell was extracted by an adaptive threshold algorithm. Secondly, SLIC(Simple Linear Iterative Clustering, SLIC) was applied to the smoothed cell image to get the superpixels of the whole cell image. Finally, based on the initial contour, we can automatically set some marks for background and foreground in a cell image full of superpixels. Then the superpixels were merged by the rule of maximal similarity. A key property of our superpixel merging is that it does not require a preset threshold, and the non-marker background regions are merged with the marked area automatically, while the non-marker superpixels are identified to avoid from being merged into background. We validate our method via the cervical cell image database and demonstrate that our method can extract the contour of cytoplasm from a single-cell cervical smear image accurately in a relatively short time.

ABSTRACT. Filling the holes of a triangular mesh has been studied for many years in the field of geometric modeling. This research is one of the reconstructing steps of a triangular mesh (or called refinement of a mesh) in order to improve the quality of a 3D triangular surface. With the same idea of hole filling in a mesh, filling in holes of 3D point clouds is still a challenge to the researchers. This paper describes a method for filling holes in an elevation surface of 3D point clouds structured in a 3D grid. The method consists of two steps. In the first step, we determine the boundary of hole. In the second step, we fill the holes based on the computation of tangent plane for each boundary point. Following clock-wise direction on the hole boundary, we compute and insert missing points on each tangent plane. This process is repeated and refined ring by ring to the inside of the hole, adapting the local curvatures of the surface. The obtained results show that the processing time of the algorithm is very fast, and the output surfaces preserve their initial shapes.

ABSTRACT. Minutiae are important features in the fingerprints matching. The effective of minutiae extraction depends greatly on the results of fingerprint enhancement. This paper proposes a novel fingerprint enhancement method for direct grayscale extracting minutiae based on the combination of Gabor filters and the Adaptive Modified Finite Radon Transform (AMFRAT) filters. First, Gabor filters are used as band-pass filters to delete the noise and clarify ridges. Next, AMFRAT filters are applied for connecting broken ridges together, filling the created holes and clarifying linear symmetry of ridges quickly. AMFRAT is the MFRAT filter, the window size of which is adaptively adjusted according to the coherence values. The small window size is used high curvature ridge areas (small coherence value), and vice versa. As the result, the ridges are the linear symmetry areas, and more suitable for direct grayscale minutiae extraction. Finally, linear symmetry filter is only used for locating minutiae in an inverse model, as “lack of linear symmetry” occurs at minutiae points. Experimental results on FVC2004 databases DB4, set A shows that the proposed method is capable of improving the goodness index (GI).

ABSTRACT. We can observe the rapid development of the spoken human-machine interface, thanks to the big progress in automatic recognition and text to speech technologies. Especially, the mobile virtual assistants are becoming more popular than ever. Recognizing user’s intent in interaction is the biggest problem while the system is designed. It is attracting the attention of researchers. Recognizing utterance’s speech act in an automated manner is important to reveal user’s intention such as informing, asking, requesting and expressing emotion, which can provide useful indicators to improve the performance of human-machine interaction. Automatic speech act classification has been studied in many languages such as English, Chinese, Slovakian, Arabic, but not yet in Vietnamese. This paper presents a speech act taxonomy suitable for Vietnamese utterances in mobile voice interaction. We also build a speech act classifier that is robust, compact and can work well on mobiles. This classifier has been applied in our mobile virtual assistant for Vietnamese(VAV).

ABSTRACT. Speech classification is one of the most vital problems in speech processing as well as spoken word recognition. Although, there have been many studies on the classification of speech signals, but the results are still limited on both accuracy and the size of the vocabulary. When classifying a huge volumes vocabulary, the speech classification becomes more and more difficult. Today, there are some frameworks that allow working with big data. One of these is a Data Mining utility, it can perform supervised classification procedures on very large amounts of data, usually named as big data, on a distributed infrastructure by using the MapReduce framework of Hadoop cluster. This tool has four classification approaches implemented. These are Random Forest, Naïve Bayes, Decision Trees and Support Vector Machines (SVM). All these approaches require input data having the same size, so the input data must be quantized before using. This leads to decrease the accuracy in the classification stage. In this paper, we propose an implementation of Local Naïve Bayes Nearest neighbor based on Hadoop framework, this allow input data with different size and work well with huge training data

ABSTRACT. We propose a method for detecting prominences in a Japanese presentation. A prominence is unclearly defined in Japanese, because it is covered only in phonetics and Japanese language education. Thus we describe the literature of phonetics and Japanese language education with actual data, and propose using the accent component as a feature for detecting the prominence. The accent component is a parameter of F0 (Fundamental Frequency), which represents the pitch accent in a word, and which is used in speech synthesis and speech analysis. In an evaluation experiment of the intensity of the accent component and its delta features, the detection accuracy was 0.80, higher than achieved in an experiment of prominence feature detection in English.