This subject uses the association rule-based recommender method to build the recommender system (MedRES – Medicine Recommender System) for the process of creating a prescription. The MedRES was built for the purpose of helping physicians choose the appropriate medicinal products to prescribe in the shortest period of time. The MedRES helps the young doctors with less experience can learn the prescribing method from other doctors. From the data set of prescription transactions, this research applies the Apriori mining algorithm to build a model on each disease. This model is a set of association rules, which represents the relationship between the drug's active ingredients.The MedRES system has good suggestions in most cases even for new doctors. The MedRES was evaluated at 2 phases: the model construction phase (the models were evaluated by experts) as well as the phase of evaluating the system after construction. In the phase after construction, the system was evaluated by using the Given-N method with the indices recall, precision and F1.

Automatic form recognition is a problem attracting attention and investment of many researchers around the word. It plays an important role in data inputting and processing in practice. The study, “Handwriting scores recognition on transcripts” isderived from the actual need of Department of Academic Affairs and the other faculties at Tay Do University. This research is carried out through many phases such as noise reducing, de-skewing, rebuilding the table, locating columns in table, and recognizing handwrittingby using Gist and SVM. Methodology for researching is to combine theory and practice for commenting and appraising the result. The research has given some highlights: a) positioning and building tables can be implemented without using any learning machines, without knowing about the logical position of the table in the form; b) locating any column as well as cell in the table without marking or color accents on the form is easily;c) the cost for form handling is low. The "Handwriting scores recognition on transcripts" system is developed successfully from the research. It allows loading the transcript from the list of transcripts image, recognizing and exporting the result to any file with the accuracy up to 97.30% on 187 transcripts.

Human posture recognition classifies a posture captured by a camera into  pre-defined postures such as stand, sit, lay. One person will do a posture in front of the camera and the system will recognize what the posture is. This paper presents the ability to recognize 20 human postures with data provided by Kinect. The advantage of using skeleton data provided by Kinect is that the result of posture recognition is invariant to the change of light condition or noise of the picture. This paper also proposes 4 feature extraction methods from the data. After that, this data will be trained by support vector machine (SVM) model. The experiments showed that the accuracy of human posture recognition is above 98%.

Mining the frequent item-set with time factor from transaction database requires much execcution time. Thus, this paper presents an approach of tree structure known as FS-Tree (Frequent Sequence Tree) to store event sequences with appearance time. From there, we propose an algorithm known as FS-Alg (Frequent Sequence Algorithm) to mine the  frequent sequence from FS-Tree. To illustrate the effectiveness of the algorithm, the paper compares it with the existing approach of TSET-Miner.

In this paper, we present the support vector machines algorithm using the stochastic gradient descent for classifying very large datasets. To reach the sparsity in the solution, the support vector machines algorithm uses the hinge loss in classification tasks. Thus, the direct optimization using the stochastic gradient descent is difficult due to the differentiation of the hinge loss. Our proposal is to substitute the hinge loss used in the problem formula of the support vector machines algorithm by the smooth ones to improve the convergence rate of the stochastic gradient descent. The numerical test results on two large textual datasets (RCV1, twitter) show that our approach is more efficient than the usual hinge loss.

A low-cost, real-time methodology for embedded devices based on well-known open source kernel - freeRTOS is presented in this study. The real-time algorithm designed consists of three main steps.Firstly, algorithm is designed and evaluated by utilizing Matlab/Simulink toolboxes. Secondly, the generic embedded C code is generated by Matlab program. Finally, freeRTOS Tasks code is utilized based on the generated C code to build and run on an embedded target. This real-time algorithm is demonstrated on a two-wheeled self-balancing robot which is employed a fuzzy PID self-tuning controller. The designed controller is executed on famous ARM Cortex M4 core microcontroller STM32F407VTG. The experimental results show that algorithm designed operated well on embedded system. The tracking position and rotation angle response in maximum delay 1.5 seconds which is fast enough while stabilizing the two-wheeled at upright. The real-time system designed is a low cost methodology and suitable for fresh embedded system designers.

This paper presents the design of a wireless-sensor-network based system for monitoring and managing on guard patrols. The aim of the project is to build an electronic system for managing guard patrols and assisting guardians in their works. The designed system is built based on MSP430G2553 chips, an Arduino Mega 2560 board, and 433-MHz wireless RF transceiver modules. By using the wireless sensor network architecture, the designed system has the ability to automatically monitor the present of guardians within a pre-defined area. The system also allows guardians to send alarm messages to the central processing unit in order to get assistances in handling urgent situations. The recorded data on guard patrols is gathered and stored on Google Spreadsheets. This allows managers to gain access to the system database from anywhere on any platform with Internet connections. The designed system is expected to be suitable for use in companies and factories.

We present, in this paper, a novel method for sign language recognition. From data acquired with Kinect camera, features of hand movement are extracted. We also propose a new feature to describe hand movement. The feature is computed by dividing the orbit of hand movement into k segments. For each segment, we compute the orientation histogram. The feature is hence independent to length of orbit. Moreover, to improve the discrimant power we also extract the visual information of hand shape with GIST feature. These features are then used to train a recognition model with support vector machines. The experimentations are realized with 280 samples collected from 5 students in Can Tho Disabled Children School. The numerical results show that the proposed method gives an 90% in term of accuracy.

Synthetic Aperture Radar (SAR) can provide various advantages, especially its function as an Earth surface observation sensor, it is becoming a very important technology in the remote sensing field. Currently, the demand for employing a SAR sensor onboard a small and low-cost satellite is highly increasing. Therefore, this paper presents research on the relationship between the parameters of the SAR system to design the SAR system on small satellites at low-cost. The paper also presents the advantages and disadvantages while choosing the operational frequency and suitable small size antenna for operational frequency. Moreover, with a small SAR, high quality of image can be achieved through the implementation of transmitter with high duty cycle.

In this paper, we propose to construct a driver drowsiness detection system using computer vision methods. A camera is used to observe driver’s face. The system will alert the driver when he had dozed off based on eyes-closed state as well as the number of eyes blinking. In this work, we use two methods to detect the eyes-closed state of the driver: distance between eye and brow; curvature of the eyelids. The first method was proposed in some previous studies while the second is novel. We develop an algorithm to determine the eyes blinking through three consecutive frames. Theexperiment on a group of Vietnamese peopleshows that the system accuracy is about 93.1%.

Recommender Systems (RS) are widely used in many fields such as e-commerce, entertainment, education, etc. The purpose of RS is to predict user preferences/behaviors/etc. Based on the prediction results, the system can recommend appropriate items to the users. This study proposes a new approach for rating prediction in RS. This approach is a compound of a voting rule and a user preference threshold to predict the rating of the user. This approach is quite simple and easy to implement but it is effective. The experimental results on standard data sets in RS show that the proposed approach performs much faster than the well-known collaborative filtering approach while its accuracy is also improved in most of the cases. Thus, this could be a promising approach for rating prediction in recommender systems.

Approaches to detecting planthopper in images can be combined with light trap systems to automate the counting of planthoppers in the light-trap images, which is currently done manually. Application of this combination would bring strong benefits to framers in monitoring the insects to control their crops. One of the most important factors that affect on the detection accuracy is the appearance of other insects in the images such as butterfly, beetles, etc. In this paper, a new approach to removing these noises will be proposed. This approach is used to removing these noises in planthopper images based on size and color of planthoppers. The experimental results show that our approach can remove the noise and help detect planthopper in images more effectively.

We present a new approach for mapping WordNet to SUMO as well as the SUMO Browser-online tool that can be used for browsing SUMO( Suggested Upper Merged Ontology), and its connection to the WordNet lexicon. The Browser facilitates the process of getting familiar with SUMO contents. In this paper, we also present SUMO and WordNet briefly.

This paper proposes a TSK fuzzy model for stock price forecasting based on Support vector machine for regression. By uniformly satisfying these conditions between TSK fuzzy models and Support vector machines for regression, we can construct an algorithm to extract TSK fuzzy model from Support vector machines. This research does not give the algorithm that allows extracting TSK fuzzy model from support vector machine, but rather proposes a solution that allows optimization of extracted fuzzy model through the adjustment of ε parameter. The proposed model is combination of the SOM clustering algorithm and fm-SVM, the algorithm to extract TSK fuzzy model from Support vector machines. The effectiveness of the proposed solutions is evaluated by the experimental results and a comparison with the results of some other models.

Chess playing is an area of research in artificial intelligence. Traditional programs were built by using Minimax, Alpha-Beta with any heuristic evaluation function based on knowledge of chess players. It is difficult to design a good state evaluation function. Moreover, A traditional tree search is suitable for the games that their branch factor is low. Monte Carlo Tree Search is a novel framework, and very effective in some high branch factors such as Go. The Monte Carlo Tree Search model is combined from Tree search, Reinforcement learning, and Monte Carlo simulation. In our view, we can improve the performance of Monte Carlo Tree Search by studying how to improve the performance of reinforcement learning, or to improve the Monte Carlo simulation. This paper compacts the most efficient way of Monte Carlo Tree Search improvement and shows its efficiency based on the experimental results.

Person re-identification through non-overlapping camera system recently becomes one of the most interesting problems in automated video surveillance. Once people are detected in the scene, they are characterized by the appearance-based models. However, the background in the cropped image obtained from the human detector could still affect the model used for person re-identification. Thus, in this paper, we propose a re-identification system including two main tasks: removing the background in the image of the detected people; and reinforcing the local patches that are discriminative to different people, and reliable in retrieving the same person in different non-overlapping cameras.

The article uses the metadata as an effective tool to remedy that defect, and illustrates it by a small software. There will be many choices for every attribute. Integrity constraints are verified appropriately to the context. All of this are always described, stored and updated automatically in the metadata. From metadata, the software automatically forms various method kinds for any class. These kinds satisfy most user requirements for an information system. In each kind, the software introduces multiple signatures that users can select to match  with class structure and the reality applying such database. The solution is general so the idea and software presented here can be applied to the other platforms and contexts. Hence, the approach is effective in raising adaptation and reuse for designing and programming.

Twitter sentiment analysis aims at classifying the comment into positive or negative sentiment. In this paper, we propose to use the bag-of-words model and the Multinomial Naïve Bayes algorithm for dealing with the sentiment classification task. In the first step, raw data sets are the comments on Twitter collected following topic. It is necessary to perform the preprocessing task, including the special characters of Twitter, continuously repeatable characters, acronyms, slang, emoticons, WordNet, and representation in Bow model. Preprocessing stage provides the large dimensional datasets in which almost values (about 99%) are zero. And then, the data set is stored in the LibSVM format (dim_index: non_zero_value). This strategy is to reduce the memory complexity and also require our new implementation of Multinomial Naïve Bayes (MNB) for dealing with the new data format. Theexperimental results on the data sets show that our implementation of Multinomial Naïve Bayes (MNB) algorithm is very simple and accurate.

Congestion detection and mitigation on wireless sensor networks has widely been pursued by the research community. Generally, wireless sensor network is deployed in dense or spare mode which has a specified strategy to detect and avoid congestion. However, the cost of implementation a wireless network is too huge so that the congestion verification should be done on the simulation before deployment in reality. We address this problem by model checking approach. Wireless sensor network is modeled by Petri nets and their properties are verified by model checking. Moreover, we use the component-based abstraction technique to abstract the unused components when verifying congestion. Our ideas have fully been implemented as a tool known as WSN-PN, on which the experimental results proved the significant improvement on verification process.

In this paper, we proposed computer vision and machine learning algorithms for an automatic road-sign detection and recognition system using HOG feature and Neural networks. Our system is able to detect and recognize almost road sign categories such as prohibition, danger, warning and information which are not overlapped. The experiments are carried out on the dataset of 31 video files. The average time to detect and identify the road signs on a frame image is approximately 0.021 seconds when using the classification model with the MLP neural network model, and approximately 0.099 seconds when using the SVM classification model. The accuracy rate for road sign identification is about 94% for both models.

Brown Planthopper (BPH) light-traps for observing and collecting infield data play the important role in pest management. Currently, these traps prove effective evidences in plant protection; however, some tasks relating to them require manual handling. Automatic mechanism light-traps are expected to replace the old ones to bring great supports for BPH prevention. This article proposes a new model of automatic sensor-based light-trap to continuously collect data. These new light-traps are adapted to form a wireless network so that they can automatically cooperate with each other. Collected data is analyzed constantly to provide instant warnings in case of high densities of BPH population. Periodically, all the data are aggregated at the control center for pest observing and predicting missions.

Minimum routing cost spanning tree problem - MRCST is a graph optimization problem that has many applications in network communication design and bioinformatics. The MRCST problem is proved to be NP-hard. Most graphs in practical are sparse graphs while the most effective algorithm solving MRCST on sparse graph is not really effective - especially with sparse graphs in large size. This paper proposes a new algorithm called HCST to solve the MRCST problem for sparse graphs. The experimental results in sparse graphs taken from the experimental data system benchmark for optimal spanning tree problems show that the quality of HCST algorithm is equal or better than that of the best algorithms known to solve MRCST in time comparison. This is also the first publication which presents theexperimental results from solving MRCST for 20 large size sparse graphs instances.

In this study, we propose a solution to build a semi-automatic consultancy system (a semi-automatic question-answering system) using mobile/Internet networks and machine learning approaches. To build the system, at first, we need to build modules for sending and receiving SMS/email messages. These modules are important for pupils to send their questions that need to be consulted. Next, a message classification module is built using a combination of machine learning method (e.g., SVM) and text processing technologies. Finally, a whole web-based system is conducted to integrate these modules. The initial results show that the system can classify the questions at 82.33% of accuracy, thus, the proposed approach is feasible.

The next generation sequencing (NGS) technologies are capable of producing low-cost data on a giga base-pairs scale in a single run, which usually includes millions of sequencing reads. This revolution allows launching many genome sequencing and re-sequencing projects for various biological applications, such as detection single-nucleotide polymorphism, and assessment of biodiversity. DNA Metabarcoding provides a door to identify the species in a large biological sequence dataset. Chloroplast genome is used as a genetic characteristic to identify species of plants. However, the traditional method to determine chloroplast genome sequence must use a sequence reference. In this paper, we propose a new approach to construct chloroplast genome sequences from raw data without using a reference sequence. To evaluate our approach, we compare the experimental result with four reference chloroplast genome sequences which were determined by biologists. The results show that the chloroplast genome sequences established by our approach are the same as the chloroplast reference sequences.

The Mekong Delta is predicted as a region that will be affected by the global climate changes. Especially, sea level rise will affect the agricultural land use in the coastal areas of the Mekong Delta. In this study, a macro model has been developed for land-use change simulation using Markov - Cellular Automata in combination with land suitability analysis method based on GAMA - an Agent Based Modeling platform. This method has resulted a model for simulating land-use change of eight provinces along the coast of the Mekong Delta from 2000 to 2008, verifying the simulation results with the land-use map in 2008 of the provinces. Base on this model, we have built the scenarios for simulation land-use in 2030 and in 2050 under the effect of sea level rise. The model could be used as an effective way for testing different land-use planning scenarios.

The rice crop constantly changes and differently distributes according to each region. Cropping calendar always changes so it is difficult for agricultural management, and rice pest early warning. The study was aimedat delineating the current and change of rice growing stages in the An Giang province by using MODIS image (MOD09Q1) at 250m spatial resolution, 8-day interval from 2011 to 2013. The results showed that MODIS image has ability to monitor the changes of rice growing stages as the basic for early warning, mainly for the large rice field. Despite the low resolution image, it can be used to monitor the current status of rice crop and pest warning in large field area and sowing simultaneously.