Fall is the most common cause of injury for elderly people. It does not only lead to physical injuries such as broken hip or head trauma, but also causes some psychological problems. However, early fall detection can help to reduce fall’s consequences. Therefore, in this paper, an approach is proposed to detect elderly fall based on accelerometer data. The fall detection model is constructed using the long short-term memory deep learning architecture. A long short-term memory with 64 hidden units is used to train the detection model. The experimental result shows that this approach is suited to detect falls of the elderly with 93.9% of accuracy.

Content-based image search has been concerned recently. This search method helps to overcome shortcomings of current meta-data-based search method, which is sensitive to the meta-data enclosed with images. In this paper, a content-based image search system is developed based on the convolutional neural network deep learning model. In addition, the search system is also combined with semenatic search technique that enables the improvement of the search result. The semance searching capacity bases  on a domain-ontology that describes semantic relationships among image topics. The experimental result shows that the accuracy of the convolutional neural network classification model on the test set is 85.75%. Moreover, the semantic search is helpful to widen and improve the search result significantly, particularly in the case that the searching keywords is ambigous or unclear.

This paper presented a model using object tracking techniques to categorize the activities of customers in the supermarket. Then, the number of customers, who were interested in the booth, were determined to evaluate the display efficiency. With the image obtained from the surveillance camera, the system can identify most of the objects entering the observation area, tracking them to obtain the trajectory and time of observation. Trajectory was segmented, and representative coordinates were used, thus using a support vector learning algorithm to classify customer activity including booth attendance and drop-in options or other activities. Also, this article proposed the improvements of the speed of object tracking algorithms in the case of tracking multiple objects at the same time. Experimentally, it found that the proposed speed improvements were significantly effective, averaging 2.8 times higher than the original, while accuracy was not changed. Data for detecting was collected from internet sources and surveillance camera data located at a large supermarket in Soc Trang province.

Deep learning is a machine learning method that has been studied and used extensively in recent years, opening up new directions for problems such as image processing, speech processing, and natural language processing, etc. This article focuses on the use of deep learning for automatic text summarization for Vietnamese. Previous approaches such as statistics, machine learning, language analysis, etc. have been successful at different levels and purposes. In this paper, the Word2vec model was used to extract the specific characteristics of Vietnamese text for the Sequence to Sequence with Attention model to produce a sequence of words. Finally, the results were re-selected using the Beam Search algorithm, and a summary sentence was generated. The accuracy of the model was estimated using the ROUGE method on a dataset of over twenty-seven million words collected from newspapers in the country. The result was the summary statement reflecting the text content. Although the results were not high yet, the model has successfully solved the problem, and the dataset needs improving to enhance the efficiency of the model.

The need for application integration is increasing, especially when organizations shift to digital economy and e-government. Enterprise servive bus is the most advantage application integration model today. Choosing an enterprise service bus solution that is suited to not only the current but also the long-term needs of an organization is a big challenge. This paper reviews several recently related research approaches and presents the results of evaluating and selecting an appropriate open-source software to build an enterprise service bus for An Giang province in the context of e-government building.

Augmented reality is a technology that combines digital information and real world in real time, input data is recorded through the camera of devices such as phone, laptop, etc. The information is usually enhanced on 3D objects, video, audio, etc. This paper presented the construction of tic-tac-toe game with human and computer players applied augmented reality technology. Image processing techniques and Hough transformation were used to detect 4 straight lines of the checkerboard, from which 9 squares were extracted. To identify the ‘X’ that the player hits, the cascade classifiers were used, each cascade classifier is a Adaboost algorithm. The experiments showed that the accuracy of ‘X’ mark recognition is above 98%. As for the problem of auto-chess, the Alpha – Beta pruning algorithm was applied.

The Mekong Delta in southern Vietnam is facing climate change and sea level rise. A solution is to quickly and accurately create new high-quality rice varieties that boost yield and adapt well to biological and non-biological factors, especially well-adapt to current harsh conditions. Since 1976, Can Tho University has collected and stored most of traditional seasonal rice varieties of the Mekong Delta. At the moment, Mekong Delta Development and Research Institute of Can Tho University has stored more than 2,000 rice variety samples. They are valuable gene resources that can be used for preserving, exploiting, employing, and creating rice varieties. However, it is possible that there are similarities in these 2,000 samples for some rice varieties. In this paper, clustering techniques are used to create tools for rice variety experts to (i) identify similar samples and (ii) analyze their similarity coefficients.

Nowadays, information technology is widely applied in agriculture - the most developed field in Viet Nam. Among these applications, the detection and recognition of pests system using handle image technique and computer vision have been attracted by many researchers. In this paper, the detection and recognition pests are resolved through two main phases: (1) detect possible areas that are pests; (2) identify the pests from the possible areas detected. In the first phase, segment method is used to detect possible areas. Binary segment and contour detection method is used to get and hightlight related objects in this phase. In the second phase, some colour features and shape features are extracted from images. Then, combined with extracted features, support vector machines are built to classify the image areas which are found in the previous phase. Classification models are trained to recognize four grapefruit leaf pests. The training results are over 99.5% for each model. The experimental result over 500 images is 99.2%. These results show that the proposed method achieves promising results and can be applied to identify the pests in reality.

This paper proposes a detection scheme of duplicate bug reports in open-source projects based on the class information associated with centroid to enhance the detection performance. This method is extended from the previous one which used only centroid method without considering the effects of both inner and inter class. Besides, this method also improved the use of normalized cosine previously for identifying the similarity between two bug reports by denormalized cosine. The effectiveness of this method is verified in an empirical study with three open-source projects, SVN, Argo UML, and Apache. The experimental results show that this method outperforms other detection schemes by about 10% in all cases.

Association rule mining, one of the most important and well-researched techniques of data mining. Mining frequent itemsets are one of the most fundamental problems and most time-consuming in association rule mining. Most of the algorithms in literature used to find frequent itemsets satisfying single minimum support threshold. In practice, frequentcy of each item reflects the nature and role of items in transactional databases. This paper proposes an efficient mining parallel algorithm for frequent itemsets with multiple minimum support thresholds (a different minimum item support for each item) on Multiple-core Processors. Proposed algorithm easily extends on distributed computing systems as Hadoop, Spark. Finally, result experiments presented on both synthetic and real-life datasets show the better proposed algorithm than the existing algorithms.

In the last few years, e-learning becomes an emergent learning method that several institutions in Vietnam have deployed at their organizations including Can Tho University (CTU). Until January 2017, the e-learning system of CTU (called Dokeos) serves for more than 950 lecturers and 50,000 students with more than 600 courses created to support the teaching of the lecturers. This study introduces e-learning and popular models in e-learning, which CTU is applying as a case study. The results of e-learning application in CTU indicated that it has been becoming a new channel for effective support in educating through credit system, contributing to training quality improvement of the university.

This paper introduces two cable-direct-driven robot (CDDR) manipulator structures which are a planar 3-cable CDDR and a planar 4-cable CDDR. Besides, a program is built to simulate kinematics models of these manipulator architectures. The simulation results show that the 4-cable CDDR requires less cable tensions and thus less energy compared to the 3-cable CDDR in performing the same simulated task. The results also highlight the possibility and reliability of these CDDRs.

Recommender system is considered one of the most effective solutions that can cope with information explosion due to the rapid development of Internet services and is widely applied in many fields. However, to design a recommender system can meet the needs of users, the selection of suitable models for the recommender system and choosing the appropriate value of parameters for the model are always big challenges of designers. This study proposes solutions to choose models and value of parameters suitable for specific collaborative filtering recommender systems. To evaluate the proposed solutions, experiments on three standard datasets of MovieLens, MSWeb, and Jester5k are conducted. Experimental results show that the proposed solutions can assist designers and researchers to quickly identify model and the value parameters model for their specific collaborative filtering recommender systems.

This study was conducted to apply geographic information system (GIS) technology and algorithms to assess land use fluctuation and land demand forecasts for socio-economic development, a case study in Ca Mau city. The GIS method and Markov chains were mainly used in this study. The data were aggregated, analyzed, and evaluated by descriptive statistics method. The results showed that the land use change in the period of 2005-2015 up to 54.2% of total area. The demand for land as forecasted by the Markov chain indicated that the area of agricultural land will reduce for conversion to residential, specialized and aquacultural land. This research has demonstrated the supporting of the GIS technology and Markov chain in the decision making and sustainable planning of land resources.

Application knowledge representation for intelligent systems is a development trend in education, especially in science technology engineering and math education. In the mathematical foundation of higher education, linear algebra is an important course. This course includes the knowledge about matrices, linear equations systems, and vector spaces. In this paper, a method for representing the knowledge domain about linear algebra is proposed. It includes three sub-domains: matrices, linear equations systems, and vector spaces. Each domain is represented by model of computational objects knowledge base. These sub-domains have been researched to combine their knowledge for solving the classes of problems in linear algebra. Based on this knowledge base, an intelligent problem solver for this course in technical universities has been built. This program can solve common exercises. Its solutions are readable, step-by-step, and alike human method.

Ontology is an advanced knowledge representation formalism. It allows reusing and sharing vocabularies between applications and plays an important role in Semantic Web. However, ontology development is complicated and time-consuming. Therefore, in this paper, an approach to constructing lightweight ontology from glossary and the WordNet was proposed. This approach based on linguistics techniques such as regular expression and Link Grammar. The experiment on the Internet Movie Database glossary showed a promising result that the proposed approach produced an ontology with more than 600 concepts and 200 relationships. However, the results still existed some limitations that required further improvements.

Cattle behaviour patterns provide significant information about cattle health. Therefore, early behaviour recognition may help breeders be aware of cattle health problems promptly to have appropriate treatment to reduce negative impact. In this paper, an approach to cow behaviour recognition based on accelerated data will be proposed. The behaviour recognition model is built using random forest algorithm. This study focuses on four popular behaviours, i.e. walking, standing, eating (grazing), and lying. The model is validated using a real cow activity datatset. The overall classification result of the model is about 95% of accuracy. The comparison on the classification result with other recent approaches is provided. It is shown that the proposed approach in this paper is promising, and it can be used for developing cow behavior recognition application.

This paper proposes a hybrid recommendation model based on statistical implicative measures to suggest a list of top N items to an active user. The proposed model is built on two sub-models: the user-based collaborative filtering model and the association rule based model. The hybrid recommendation model is compared to its sub-models and some existing models such as latent factor model, popular model, and user-based collaborative filtering using Cosine on two datasets MSWeb and DKHP. The experimental results show that the performance of the proposed model is better than the compared models.

These days, digital music storage systems (DMSS) in Vietnam usually arrange pieces of music according to the composer’s name and the song’s title, whereas listeners need to search for songs based on genres and contents. This increases the demand for categorizing songs in accordance with genres in DMSS, which enables listeners to search for the most wanted music. However, with a large number of songs collected, the way to classify them for easy management becomes a challenge for all DMSS. Therefore, it is necessary to build up an automatic sorting system. This paper suggests a new method of extracting specific timbral disposition including timbral texture, rhythmic content by using wavelet convert. Thanks to such distinctive features, KNN and SVM methods are utilized to identify types of music files. This study is conducted on four types of music: Bolero, Cai luong (reformed theatre), Cheo (classical theatre) and Hat Boi (traditional opera). The findings show that the reliability is up to 93.75% and 94% corresponding to KNN and SVM on the timbral texture. Moreover, these suggested methods are simple, effective, speedy, and suitable for Vienamese music sorting systems today.

This paper proposes the construction of a recommender system to predict users’ preferences based on matrix factorization techniques. Because of the changes of users’ preferences time by time, to achieve more accurate result, exponential smoothing is integrated into the matrix factorization model by utilizing tensor factorization. This usage aims at exploiting and taking  advantage of information about the time and the order of users’ giving feedbacks. The model is tested relied on the datasets in suggestion and evaluation using the root mean squared error. The experimental results demonstrate fairly good performance of the proposed method.

This paper presents a revised version of a cluster number estimation algorithm for big datasets. This algorithm was designed to work on a standard personal computer. This is an improvemennt of the Cell-MST-Based cluster number estimation algorithm by appying weighted distance instead of using the Euclidean distance. This new algorithm was named Weighted-Cell-MST-based cluster number estimation algorithm. This revised version can provide more stable results compared to its former version when testing the same datasets in the same environment.

Blast disease is caused by the Pyricularia oryzae fungus and has been recorded in more than 80 rice-producing countries around the world, and the disease is progressively more complex, causing many difficulties for farmers. From previous studies on rice blast disease, six important factors including rice variety, density, temperature, moisture, leaf color (protein), and lesion status have been found to have significant influence on the pathogenesis of diseases. Today, with the rapid development of internet networks, mobile devices, etc., most of the farmers own mobile phones. In this study, the content-based of recommender method is used to build the mobile application “BLASTREC” that supports farmers in blast prevention. The software BLASTREC functions Android operating system based on two Naive Bayes and Decision Tree classification algorithms. Experimental results show that the accuracy of two algorithms is more than 90%. The experiment data on blast in Trung An area, Thot Not district, Can Tho city combines with agricultural experts’ opinion to provide farmers with appropriate treatment.

Stroke (Brain attack) is not only one of the two reasons leading human death over the world but also the most popular and dangerous cerebrovascular disease. In Vietnam, the lack of the specialized equipment as well as the force of qualified experts has become the massive problem for the accurate diagnosis as well as the efficient and well-timed treatment of stroke, especially intracerebral hemorrhage, the acute type of stroke. Based on analyzing and giving the solutions for the challenges, a service-oriented architecture for the big data-driven knowledge management system on medical intracerebral hemorrhage images was proposed. The suggested architecture provides the capability to develop the intracerebral hemorrhage knowledge systematically, consisting of the knowledge exploration and the knowledge exploitation. As a result, it can contribute to the timely and effective support in the treatment of intracerebral hemorrhage. Besides, the architecture adapts to the modern knowledge service modeling tendency. According to this trend, the knowledge management system can be expended, shared and integrated with more knowledge contributed from specialists, doctors, hospitals, and research institutes.