New Trends in Computer Technologies and Applications

Nowadays ILP processors can’t analyze the semantic information of instruction thread to change instruction series automatically for increasing ILP degree. High performance required programs such as image processing or machine learning contain a lot of loop structure. Loop structure will be bounded with the instruction number of one basic block. That cause processors are hard to enhance the computing efficiency. The characteristics of the loop structure in the program are as follows: (1) Instruction will be fetched from cache and be decoded repeatedly. (2) The issued instructions are bounded by the loop body. (3) There is data dependence between iterations. These factors will get worse the poor ILP in the loop codes. In this paper, we propose an architecture called semantic-based dynamic loop unrolling mechanism. The proposed architecture can buffer the instruction series of nested loop, unroll it automatically by analyzing the instruction flow to find the loop body with the semantic of loop instructions, store them to the instruction buffer, and dispatch them to target the processor cores. The proposed architecture consists of three units: loop detect unit (LDU), unrolling control unit (UCU) and loop unrolling unit (LUU). LDU will parse the semantic of instructions to find the closed interval of the loop body instructions. UCU will control LUU in the whole process. LUU will unroll the loop based on the information collected by LDU. Loop controller will handle the complementation overhead for branch miss prediction and the loop finish-up codes. The verifications use ARM instructions generated by Keil $$ \mu $$ Vision5 compiler. The results show that eliminating iteration dependence can improve ILP by 140% to 180%.

The local-dimming backlight has recently been presented for use in LCD TVs. However, the image resolution is low, particularly at weak edges. In this work, a local-dimming backlight is developed to improve the image contrast and reduce power dissipation. The algorithm enhances low-level edge information to improve the perceived image resolution. Based on the algorithm, a 42-in. backlight module with white LED (Light-Emitting Diode) devices was driven by a local dimming control core. The block-wise register approach substantially reduced the number of required line-buffers and shortened the latency time. The measurements made in the laboratory indicate that the backlight system reduces power dissipation by an average of 48 percents and exhibits no visible distortion compared relative to the fixed backlighting system. The system was successfully demonstrated in a 42-in. LCD TV, and the contrast ratio was greatly improved by a factor of 100.

This study, we present a novel robot localization system based on Zigbee nodes. The Zig-Bee locator can provide relative indoor position information. Based on the Zigbee locator, the computer calculates the sensing data and its results are sent to the micro-controller to control motors, to enable make robot walking on the middle of the passageway. The system is successfully implemented and demonstrated in real environment.

In recent years, there is a rising interest in aiding people with computer technologies. One of the possible research ways is the use of speech interface for the low resource device such as mobile phones or PDAs. In this paper, we proposed a systematic way to construct a speech interface for the mobile device such as an Android phone. First, an automatic speech recognition system is built based on the Hidden Markov Model (HMM) framework. In order to understand the utterance from user in daily life, the spontaneous speech recognition is implemented. For speech generation, a personalized speech synthesis system is also included for the proposed system. Experimental results have shown that the performance of the speech recognition and synthesis achieve good performance. Therefore, the proposed speech interface could be applied to other system such as computer vision system in order to help people to perceive the environment and interact with other people in daily life.

In response to the concept of Smart City and Internet of Things (IoT) being prevailed in recent years, networking requirements for various devices have increased significantly, however, many related problems have arisen. To solve the problems and shortcomings of the existing IoT wireless networking technologies, for example, the data rate of ZigBee is slow and the wireless networking technologies which uses star topology like NB-IoT and LoRa is rely on the signal of base station, if the node is at communication deadzone, it will not be able to successfully transmit data. Therefore, this paper proposes an Adaptive Linked-List Mechanism For Wi-Fi Wireless Network, this algorithm takes advantage of the low-collision of the linked-list architecture and solves the problem that multiple nodes cannot be started at the same time, moreover, the nodes are able to use signal strength as the basis for constructing a linked-list network that ensure that the nodes within the signal coverage can be stably connected, and the entire system will not be interrupted due to a problem with a few nodes. The algorithm also modifies the linked-list network construction and broken node recovery process, and adds the packet signal strength queue mechanism, so that multiple nodes can quickly construct the network chain. Furthermore, we design a Wi-Fi control board which is based on CC3220MODA and implement our algorithm on it, finally, after testing in different environment, we offer two objective function for two important arguments. We hope to take advantage of this wireless linked-list network in data collection network in the future, and then solve the multi-device networking requirements, and build a stable, reliable and high-speed data collection network for IoT.

The concept of smart grid has been proposed for years. Many countries, such as United State, England and Japan, have been replacing traditional electric meters with smart electric meters in recent years. There are lots of communication methods used in smart grid, power line communication (PLC) is an important one among them. G3-PLC is a widely-used specification for long-distance PLC, however, PLC is sensitive to old power lines and the interference caused by large electric current flowing through the power line. Moreover, although G3-PLC has stable performance of communication, the AODV routing protocol and the complex startup procedures results in taking long time for G3-PLC devices to finish the whole startup procedures.To reduce the time to finish the startup procedures in G3-PLC, in this paper, a tree topology algorithm is proposed. In the tree topology algorithm, a simple startup procedure is provided, by setting up the parent-child relationship between nodes, the routes for nodes in networks are simplified. Furthermore, a maintain procedure is also provide in the algorithm, in the maintain procedure, the nodes in networks can use Check Alive mechanism to check the connection between their parent nodes and child nodes, when a broken connection is found by a node, the node will use Recovery mechanism to rescue the isolated nodes.In this paper, the feasibility of the tree topology algorithm is verified by NS-3 platform, and suggestions of the suitable value of the parameters in the algorithm are proposed. In addition, the algorithm is implemented on Atmel SAM4CP16C evolution kits, the measurement results show that the time for PLC devices to finish the startup procedures of the tree topology algorithm is at least 6.7 times less than G3-PLC specification.

Advanced Persistent Threat (APT) has become one of the most complicated and intractable cyber attack over the last decade. As APT attacks are conducted through series of actions that comprise social engineering, phishing, command and control servers, and remote desktop control, conventional anti-virus mechanisms become insufficient because they were designed to cope with traditional stand-alone malware attacks. Furthermore, data transmission from the compromised network to the APT actors is usually well disguised and embedded in normal transmission, exacerbating the detection of APT attacks to the point that even major anti-virus firms are not sure about the ratio of discovered APT attacks against real attacks. To make things worse, APT actors tend to be well-organized and potentially government-funded groups of hackers and professionals who are capable of developing and maintaining malware specifically made for their own purposes and interpret the stolen data. While most efforts in defending against APT attacks focus on related technologies, this research argues the importance of constructing a holistic understanding by analyzing the behaviors and changes of ATP attacks and actors. This research aims to understand the evolution of technologies and malware on the one hand and the behavioral changes of attacking groups. By doing so, this research is expected to contribute to constructing a clearer roadmap of APT attacks and actors that cyber security providers can use as reference.

In recent years, many studies on peer-to-peer (P2P) botnet detection have exhibited the excellent detection precision on synthetic logs collected from the testbed. However, most of them do not evaluate their effectiveness on real traffic. In this paper, we use our BotCluster to analyze real traffic from April 2nd to April 15th, 2017, collected as Netflow format, with three time-scopes for detecting P2P botnet activities in two campuses (National Cheng Kung University (NCKU) and National Chung Cheng University (CCU)). Three time-scopes including single-day, three-day, and weekly observation period applied to the same traffic logs for revealing the influence of the observation period on P2P botnet detection. The experiments show that with the weekly observation period, the precision can increase 10% from 84% to 94% on the combined traffic logs of two campuses.

Convolution neural network recently confirmed the high-quality reconstruction for single-image super-resolution (SR). In this paper we present a Deep Level Residual Network (DLNR), a low-memory effective neural network to reconstruct super-resolution images. This neural network also has the following characteristics. (1) Ability to perform different convolution size operations on the image which can achieve more comprehensive feature extraction effects. (2) Using residual learning to expand the depth of the network and increase the capacity of learning. (3) Taking the skill of parameter sharing between the network module to reduce the number of parameters. After the experiment, we find that DLNR can achieve 37.78 in PSNR and 0.975 in SSIM when using Manga109 as testing set for 2× SR.

For markerless indoor Augmented Reality Navigation (ARN) technology, camera pose is inevitably the fundamental argument of positioning estimation and pose estimation, and floor plane is indispensably the fiducial target of image registration. This paper proposes optical-flow-scene indoor positioning and wall-floor-boundary image registration to make ARN more precise, reliable, and instantaneous. Experimental results show both optical-flow-scene indoor positioning and wall-floor-boundary image registration have higher accuracy and less latency than conventional well-known ARN methods. On the other hand, these proposed two methods are seamlessly implemented on the handheld Android embedded platform and are smoothly verified to work well on the handheld indoor augmented reality navigation device.

Augmented Reality (AR) headsets has become a potential device as an auxiliary tool for practicing physical activities such like Tai-Chi Chuan (TCC). Although some learning systems can display the virtual coach movement in AR headsets, the playing speed cannot be adjusted appropriately just like a real coach stand next to you. In most of the learning system, the common approach is using controller to control the playback system under a specific speed. Once user want to speed up or speed down, he has to do these commands via a controller. In this work, we propose a TCC learning system which will real time detect the delay time between current action of user and virtual coach. After obtaining the delay time, our learning system will adjust speed of virtual coach automatically. With real time speed adjustment, learners can practice TCC with their own pace and virtual coach will slow down or speed up to follow learners’ movement.

Community detection in social networks is widely studied because of its importance in uncovering how people connect and interact. However, little attention has been given to community structure in Facebook public pages. In this study, we investigate the community detection problem in Facebook newsgroup pages. In particular, to deal with the diversity of user activities, we apply multi-view clustering to integrate different views, for example, likes on posts and likes on comments. In this study, we explore the community structure in Facebook public pages. The results show that our method can effectively reduce isolates and improve the quality of community structure.

Learning-based hashing has been widely employed for large-scale similarity retrieval due to its efficient computation and compressed storage. In this paper, we propose ResHash, a deep representation hash code learning approach to learning compact and discriminative binary codes. In ResHash, we assume that each semantic label has its own representation codeword and these codewords guide hash coding. The codewords are attractors that attract semantically similar images and are also repulsors that repel semantically dissimilar ones. Furthermore, ResHash jointly learns compact binary codes and discover representation codewords from data by a simple margin ranking loss, making it easily realizable and avoiding the need to hand-craft the codewords beforehand. Experimental results on standard benchmark datasets show the effectiveness of ResHash.

With the advancement of technology, image processing has become very common and widely used in the field of detection and recognition. This paper uses the digital image processing method to capture the feature of breads and recognize them. The experimental results: First detect the bread and then capture the image features. Finally, using Logistic Regression to classify them. The highest score as the recognition result. Through the above we can achieve effective and outstanding results.

In recent years, rapid growth of motorcycles enables a large number of traffic accidents. Hence, how to manage the traffic flow has been a hot topic. In this paper, we propose a method for recognizing and counting the motorcycles by integrating the support vector machine (SVM) and convolutional neural network (CNN). In this work, the CNN is first adopted to generate the implicit features, and then the SVM is trained based on the implicit features and tested for unknown images. The experimental results reveal the proposed method can achieve low error rates in counting motorcycles.

The smart transportation system is one of the most essential parts in a smart city roadmap. The smart transportation applications are equipped with CCTV to recognize a region of interest through automated object detection methods. Usually, such methods require high-complexity image classification techniques and advanced hardware specification. Therefore, the design of low-complexity automated object detection algorithms becomes an important topic in this area. A novel technique is proposed to detect a moving object from the surveillance videos based on CPU (central processing units). We use this method to determine the area of the moving object(s). Furthermore, the area will be processed through a deep convolutional nets-based image classification in GPU (graphics processing units) in order to ensure high efficiency and accuracy. It cannot only help to detect object rapidly and accurately, but also can reduce big data volume needed to be stored in smart transportation systems.

Face tracking methods are increasingly critical for many expression mapping analysis applications, along its research track, deep convolutional neural network (DCNN-based) search techniques have attracted broad interests due to their high efficiency in 3D feature points. In this paper, we focus on the problem of 3D feature point’s extraction and expression mapping using a light-weight deep convolutional neural network (LW-DCNN) search and data conversion model, respectively. Specifically, we proposed novel light-weight deep convolutional neural network for 3D feature point’s extraction to solve the great initial shape errors in regression cascaded framework and the slow processing speed in traditional CNN. Furthermore, an effective data conversion model is proposed to generate the deformation coefficient to realize the expression mapping. Extensive experiments on several benchmark image databases validate the superiority of the proposed approaches.

Three-dimensional (3D) urban reconstruction becomes increasingly crucial in many application areas, such as entertainment, urban planning, digital mapping. To achieve photorealistic 3D urban reconstruction, the detailed reconstruction of building facades is the key. Light Detection and Ranging (LiDAR) point clouds and images are the two most important data types for 3D urban reconstruction, which are complementary regarding data characteristic. LiDAR scans are sparse and noisy but contain the precise depth data, whereas images can offer the color and high-resolution data but no depth information. In recent years, an increasing number of studies show that the fusion of LiDAR point clouds and images can attain better 3D reconstruction results than a single data type. In this paper, we aim to provide a systematic review of the research in the area of the 3D facade reconstruction based on the fusion of LiDAR and images. The reviewed studies are classified by the different usage of images in the reconstruction process. We hope that this research could help future researchers have a more clear understanding of how existing studies leverage the data in LiDAR scans and images and promote more innovations in this area.

The study proposed an application to detect cryptomeria trees which had been damaged by squirrels from an aerial image. Each pixel of the aerial image is classified into damaged tree pixels or healthy pixels by super vector machine (SVM) developed in this paper. The application achieves about 82.24% true positive rate (TPR) in detecting damaged trees and about 1.30% false positive rate (FPR). It is a smart tool for monitoring smart forests.

Convolutional Neural Network (CNN) has demonstrated great efficiency in image classification tasks. In this paper, CNN is used with multi-label classification to extract features from clothing images and classify clothing types and colors. In our method, we first develop a neural network for types and colors separately then combine into a multi-output model to identify them. The experimental results show that the proposed method achieves practical performance in classify precision.

This paper examines different methods of producing real-time Agtron index outputs for coffee bean baking. The goal is to provide an optimal roasting output based on the required profile, increasing baking accuracy over the commonly used time-temperature method. Although the Agtron baking degree is based on the caramel infrared index, it is also highly correlated with color and shape information. Experimentally, a baking color was sub-divided into ten categories (grades), images were taken with a common color camera, then a deep learning convolutional neural network performed analysis. Based on the LenNet architecture and parameters, this study develops a “convolution neural network for coffee bean baking identification” and develops a time-sequential binary classification model (TSBC) based on the time-decreasing characteristics of baking. The resultant system correctly determines the baking grades.

The goal of this work is to enable users to try on clothes by photos. When users providing their own photo and photo of intended clothes, we can generate the result photo of themselves wearing the clothes. Other virtual try-on methods are focused on the front-view of the person and the clothes. Meanwhile, our method can handle front and slightly turned-view directions. The details of the clothes are clearer. In the user study, about 90% of the cases, respondents chose our results over others.

Scene recognition is one of the hallmark tasks in computer vision, as it provides rich information beyond object recognition and action recognition. It is easy to accept that scene images from the same class always include the same essential objects and relations, for example, scene images of “wedding” usually have bridegroom and bride next to him. Following this observation, we introduce a novel idea to boost the accuracy of scene recognition by mining essential scene sub-graph and learning a bi-enhanced knowledge space. The essential scene sub-graph describes the essential objects and their relations for each scene class. The learned knowledge space is bi-enhanced by global representation on the entire image and local representation on the corresponding essential scene sub-graph. Experimental results on the constructed dataset called Scene 30 demonstrate the effectiveness of our proposed method.

In this paper, an effective methodology which hybridizes a LCS finding algorithm and SimHash computation is presented for evaluating the text-similarity of articles. It reduces the time-space scale needed by the LCS algorithm by breaking the articles into word subsequences of sentences, managing and pairing them by SimHash comparisons, and reaching the goal of evaluating long-length articles rapidly, with the similar parts and similarity score of compared articles figured out exactly.

In order to find passengers’ behaviors when the passengers take buses, 456 thousand and 82 million records of electronic ticket transactions of route 151 and Taichung City Bus in 2015 are respectively analyzed in this article. There are three statistical/analytic results. First, about 5.26 million electronic ticket users received benefits from Taichung City Government’s policy for a free bus ride within 10 km with an electronic ticket; however, less than 0.5% users still used cash. Second, The passengers usually got on and off route 151 at THSR Taichung Station no matter which direction. Other bus stops for passengers usually getting on and off were T.P.C.C., Wufeng Agr. Ind. Senior High School, Wufeng, and Wufeng Post Office. Finally, on Friday and the day before holidays, many passengers changed their behaviors to take route 151 from Wufeng District to THSR Taichung Station. This change was that the passengers took another bus route to the station near the start station of route 151 to increase the probability to get on the route 151.

In order to find passengers’ behaviors when the passengers take buses, more than 82 million records of electronic ticket transactions of Taichung City Bus in 2015 and 8 inter-county/city bus routers are analyzed in this article. There are three statistical/analytic results. First, about 5.26 million electronic ticket users received benefits from Taichung City Government’s policy; however, less than 0.5% users still used cash. The situations of 8 inter-county/city bus services in adjacent counties were little similar to that of Taichung City. Second, route 208 was the major route took by most people between Taichung City and Miaoli County among the routes supported by Taichung City Government. Routes 108 and 101 were the major routes connecting Nantou County and Taichung City, and Changhua County and Taichung City, respectively. Finally, in three adjacent counties, the names of top 5 bus stops for each county are almost same, but the order of top 5 bus stops for each county are quite different.

The solar energy is getting popular due to the awareness of the environmental issues. Multiple module strings are set up in a solar-power plant to increase power production which is sold to electricity company via connected grid. Inevitably, devices can break, leading to loss of power production. To minimize the loss, it is important to be able to detect faulty devices as soon as possible for maintenance. In this paper, an approach relying on careful data pre-processing is proposed and compares with an existing approach.

To keep high-performance operations in photovoltaic system is an important task. However, solar panels are subject to pollution in the natural environment. Possible pollutions include dust in the air, feces in birds and dust from burning materials, etc., which can cause solar energy performance reduced. In this paper, the effects by manual cleaning and natural cleaning are investigated and the result is aimed to help manager to determine the timing of cleaning. The Array Ratio (RA) was used to evaluate the daily power generation performance and various conditions which can prevent from obtaining reliable RA values were addressed. Experiments were conducted to verify the proposed approach. The results showed that the first manual cleaning has a significant cleaning effect of 10.11% and the second manual cleaning effect only 2.03% since before the cleaning there were several heavy rains, resulting in natural cleaning of the plant.

Detection and analysis of anomalous communication behaviors in cellar networks are extremely important in identifying potential advertising agency or fraud users. Visual analytics benefits domain experts in this problem for its intuitiveness and friendly interactive interface in presenting and exploring large volumes of data. In this paper, we propose a visual analytics system, egoStellar, to interactively explore the communication behaviors of mobile users from an ego network perspective. Ego network is composed of a centered individual and the relationships between the ego and his/her direct contacts (alters). Based on the graph model, egoStellar presents an overall statistical view to explore the distribution of mobile users for behavior inspection, a group view to classify the users and extract features for anomalous detection and comparison, and a ego-centric view to show the interactions between an ego and the alters in details. Our system can help analysts to interactively explore the communication patterns of mobile users from egocentric perspectives. Thus, this system makes it easier for the government or operators to visually inspect the massive communication behaviors in a intuitive way to detect and analyze anomalous users. Furthermore, our design can provide the researchers a good opportunity to observe the personal communication patterns to uncover new knowledge about human social interactions. Our proposed design can be applied to other fields where network structure exists. We evaluated egoStellar with real datasets containing the anomalous users with extremely large contacts in a short time period. The results show our system is effective in identifying anomalous communication behaviors, and its front-end interactive visualizations are intuitive and useful for analysts to discover insights in data.

As the result of Social media data being unprecedentedly available, we are provided so many substantial opportunities to explore social circle from many perspectives. Many researches have deep understandings of the correlation between social relation online and that on real world. These researches make conclusions from perspectives of statistics only but they cannot explain specific reasons underlying those conclusions. Obviously we can separate an individual’s social circle into several groups such as family group, classmate group and co-worker group etc. Aimed at labeling the friend groups online, in this paper, we implement a visual analytic system to exploring correlation between communities resulted from label propagation algorithm in online social friendship network of a centric user and each friend’s offline POI distribution of such places where the centric user has also visited. To demonstrate the reasonability and utility, we give 3 cases in details based on social media data of the online friendship and offline movement information provided by Tencent (the largest social service platform in China).

Futures market has high leverage and the characteristics of over-performing returns. Hence, it always attracts lots of investors. However, three major kinds of institutional traders are more influential than individual investors in Taiwan. In this paper, a monthly predicting model for the weighted price index of the Taiwan Stock Exchange (TAIEX) was built based on correlation and regression analysis by using the following parameters: Dow Jones industrial index, NASDAQ index, M1A, M1B, M2 annual growth rate, US dollar exchange rate, economic monitoring indicator and global oil prices. Then, based on the prediction results, we define two trading strategies and apply them in trading MTX (mini Taiwan index futures). The evaluation result shows that both the two trading strategies have good returns.

Latent variable model has been widely used to extract topics from document collections, but this unsupervised mode makes it difficult to interpret the topics covered by the constructed cluster due to no artificial tags. Considering the influence of preprocessing on subsequent data mining, this study extracts news feature words from CSCP word segmentation method and named entity identification (NER). For NER, names of people, places, organizations are extracted by syntax rules and verified by Wikipedia. In terms of CSCP, the probability of continuing characters, the distribution probabilities and numbers of links before and after the character are used to constantly merge the unit words, extract the important narrative words effectively, and then conduct Unigram, Compounds and Mixture word processing. The corpus of words obtained after word processing by CSCP and NER were used as LDA prior knowledge. Finally, the topic coherence of NER and CSCP is evaluated by UMass Topic Coherence Measurement. The experimental results show that Compounds are of specific meaning; Mixtures represents its diverse scope; Unigrams and NER are relatively short, while NER can accurately represent the important features of news content, the topic is more cohesive. In terms of efficiency, NER-LDA takes the longest, while it had the highest degree of topic coherence.

Live-cell time-lapse images with particles generated by experiments are useful for observing results, even for proposing novel hypotheses. By identifying particles and cells as objects from these images and then calculating measures from them, such as the distances, they can be quantized for the relationship of particles and cells. However, this work is very time-consuming when calculating the distances among a large number of images. Hence, a very important issue will be presented here in order to accelerate the calculations. In this paper, we will propose parallel algorithms for calculating particle-cell distances, abbreviate to a PCD problem. Two parallel PCD algorithms, called PPCDOMP and PPCDCUDA, will be developed by using OpenMP and CUDA. After the experimental tests, the PPCDOMP with 16 CPU threads achieves 11.7 times by comparing with the PCD algorithm in a single thread; however, the PPCDCUDA with 256 GPU threads per thread block only achieves 3.2 times. Therefore, the PPCDOMP algorithm is suitable for analyzing live-cell time-lapse images with particles based on the shared memory environments.

The language NESL aims to facilitate GPU programming. In order to utilize the computation power of GPUs, NESL programs must be translated into efficient low-level code for execution. We propose a new translation technique. In NESL, apply-to-each is the main construct to extract parallel computation capability of GPUs. The result of apply-to-each is a sequence of elements. In traditional translation, the order of the elements in a sequence is always preserved. However, sometimes, the order need not be preserved and hence a faster method (which may not preserve the order of elements) for calculating the sequence may be employed. We propose the order analysis to determine if the order of elements in a sequence needs to be preserved. Order analysis is based on the taint analysis. In our experiments, we obtained 8.76x speedup on average.

Cloud computing benefits emerging Inter of Things (IoT) applications by providing virtualized computing platform in the cloud. However, increasing demands of low-latency services motivates the placement of computing platform on the edge of network, a new computing paradigm named fog computing. This study assumes container as virtualized computing platform and uses Kubernetes to manage and control geographically distributed containers. We consider the design and implementation of an auto-scaling scheme in this environment, which dynamically adjusts the number of application instances to strike a balance between resource usage and application performance. The key components of the implementation include a scheme to monitor load status of physical hosts, an algorithm that determines the appropriate number of application instances, and an interface to Kubernetes to perform the adjustment. Experiments have been conducted to investigate the performance of the proposed scheme. The results confirm the effectiveness of the proposed scheme in reducing application response time.

This paper proposes an internet of things system which can generates the map of people emotion around the campus. The system can identify someone’s emotion who stays in a school. Based on image retrieving, face detection and emotion recognition method, the emotion information can be gotten, store in the database and user can query the instance of emotion around the campus by the system. The query function is combined into an exist system which the authors proposed previously. The application of the system includes provide the professor to observe the students’ statement and adjust the teaching mode, activity result evaluation. Some system demonstration is also shown in the paper.

With the rapid development of the information and communication technology, the security requirement of the entrance guard control systems has significantly increased. In recent years, the rise and popularity of various portable smart devices, including traditional keys, inductive buckles, inductive magnetic disks, biometrics, and so on. It has been already ripe that the biometrics, especially the finger-print identification technology from intelligence mobile phones, intelligence watches, etc., has matured and become an important member of an entrance guard control system. This research is to enhance the entrance guard control management in the traditional rental housing industry with a new generation of intelligent ultrasonic bio-metric entrance guard control system. In addition to preventing loss of cards and copy of keys, the research can also have the feature of achieving remotely monitor of the rental situation of a rented house. Even if it is not in the local area, the security of the rental house can be immediately controlled by the system to achieve the optimization result of the unattended intelligent lease access control system. Finally, the of goal of this study is to provide a convenient and safe living environment for rental houses.

Visual data (e.g., images/videos) captured from outdoor visual devices are usually degraded by turbid media, such as haze, rain, or snow. Hence, weather conditions would usually disrupt or degrade proper functioning of vision-based applications, such as transportation systems or advanced driver assistance systems, as well as several other outdoor surveillance-based systems. To cope with these problems, removal of weather effects (or the so-called deweathering) from visual data has been critical and received much attention. Therefore, it is important to provide a preprocessing step to automatically decide the current weather condition for input visual data, and then the corresponding proper deweathering operations (e.g., removals of rain or snow) will be properly triggered accordingly. This paper presents an inception network-based weather image classification framework relying on the GoogLeNet by considering the two common weather conditions (with similar characteristics), including rain and snow, in outdoor scenes. For an input image, our method automatically classifies it into one of the two categories or none of them (e.g., sunny or others). We also evaluate the possible impact on image classification performance derived from the image preprocessing via filtering. Extensive experiments conducted on open weather image datasets with/without preprocessing are conducted to evaluate the proposed method and the feasibility has been verified.

While automated optical inspection (AOI) is an effective means to evaluate the quality of wood product, there are very few other applications of AOI in the domestic wood factories. For example, defects of holes or cavities on the surface of wood product currently are still found and filled with glue by human labor. In this paper, we propose using the laser triangulation method of 3D machine vision, to detect defects (holes or cavities) for a large area of lumber surface, and the information collected can be used in the automatic filling machine for further filling process. This method proved to be unaffected by wood surface texture, and can identify the cavity defect with a size larger than 1 mm under the speed of 3.6 m/min for wood with 1.5 m wide.

Testing the Internet of Things (IoT) solution is complex as it involves a diversification of implementation of smart objects that adopt a diverse and complex communication protocols. It is doubtful whether tests done in IoT solution have been adequately sufficient and scalable. This paper proposed a mapping of the IoT architecture to the conventional software test levels. The test levels shall provide a better view for tester to conduct tests based on different focus of the level.

Based on the sensing network to implement the fire emergency escape path planning. Unlike other planning methods, this paper uses grid technology to consider factors such as the impact of crowd density on travel time, smoke diffusion, and the upper limit of space/aisle accommodation. At the time of the fire, the evacuation path planning problem at each location is transformed into the optimization. In the model, a balanced evacuation was considered to achieve the goal of evacuation of the crowd. In the paper, the simulated annealing algorithm was used to solve the optimization model and the experimental results were verified.

In this paper, a PC/PLC integrated mixer system is proposed and implemented. The proposed system uses a PLC (Programmable Logic Controller) to control the mixing stage and two PCs for monitoring the whole process and dispatching the daily manufacturing order. One of PCs works as a database server that provides a web interface to the manufacturing manager. In the central control room, the manufacturing manager dispatches the manufacturing order, including formulas and the number of batches. These commands are sent to another PC, an industrial PC, which can communicate with the PLC. Some parameters are also sent to PLC via this IPC, e.g., the times of each step. The IPC gathers the system information and checks that all materials are met the criterion. The IPC also shows the real-time manufacturing information on the screen to inform the operators. With the integration of the advantages of PC’s networking capabilities, data storage capacity and PLC’s real-time control, the proposed system provides a new generation of industrial 4.0 automated production system which is more stable and can improve the quality of production.

GPS satellite positioning is currently a system used by people for navigation. As urban buildings become larger and their internal usages more diverse, the internal structure of the buildings turns into complex. Indoor positioning with GPS is often blocked by buildings, so that GPS is unable to accurately locate a user’s position inside buildings. Currently, dead reckoning techniques can improve this issue in indoor positioning. However, the techniques need the deployment of many sensing devices in indoor environment, and the estimated distance would still result in serious errors. In this situation, it leads to the high cost of labor, materials, and equipment. In this study, we develop a gait recognition and counting scheme based on a self-made sensor to collect spatial data and user walks. The scheme can accurately calculate the number of walking steps, and then calculate the distance by effectively reducing the distance error.

The results of this study show the college students’ personal test records, which the average index of college students’ cardiorespiratory strength (3 min) is 29.32, the average index of muscle fitness (1 min sit-ups) 37.71, the average score of softness sitting (frontal bend) 26.53, the average index of body mass index 37.71, and the average index of explosive force (established long jump) 174.63. The goals of exercise is to strengthen muscle fitness, heart and lung, endurance, and weight control. The type/project is ranked first in selecting medium-intensity activities, followed by high-impact activities such as jogging, and finally low-impact activities. Set the exercise target frequency for the selected sports to take the first place in 3 days a week. The exercise intensity is mostly due to the fact that people have diabetes mellitus and are a little tired (or load), but most of them are painful. The duration of exercise is 1 point per 5 min, the average index is 8.81, and most of them are 8 to 10 points each time. Therefore, the most frequent duration of each exercise is 40 to 50 min.

Given two sequences A and B of lengths m and n, respectively, and another constrained sequence C with length r, the constrained longest common subsequence (CLCS) problem is to find the longest common subsequence (LCS) of A and B with the constraint that C is contained as a subsequence in the answer. Based on the diagonal concept for finding the LCS length, proposed by Nakatsu et al., this paper proposes an algorithm for obtaining the CLCS length efficiently in O $$(rL(m-L))$$ time and O(mr) space, where L denotes the CLCS length. According to the experimental result, the proposed algorithm outperforms the previously CLCS algorithms.

Let T1, T2 be spanning trees in a graph G. If for any two vertices u, v of G, the paths from u to v in T1, T2 are vertex-disjoint except end vertices u and v, then T1, T2 are called two completely independent spanning trees (CISTs for short) in Pai and Chang [12] proposed an approach to recursively construct two CISTs in several hypercube-variant networks, including crossed cubes. For every kind of n-dimensional variant cube, the diameters of two CISTs for their construction are 2n − 1. In this paper, we give a new algorithm to construct two CISTs T1 and T2 in n-dimensional crossed cubes, and show that diam(T1) = diam(T2) = 2n − 2 if n ∈ {4,5}; and diam(T1) = diam(T2) = 2n − 3 if n ≥ 6 where diam(G) is the diameter of graph G.

Location-based social networks are important issues in the recent decade. In modern social networks, websites such as Twitter, Facebook, and Plurk, attempt to get the accurate address positions from their users, and try to reduce the gap between virtuality and reality. This paper mainly aims at both the interests of Internet users and their real positions. This issue is called the spatial-aware interest group query problem (SIGQP). Given a user set U with n users, a keywords set W with m words, and a spatial objects set S with s items, each of which contains one or multiple keywords. If a user checks in a certain spatial object, it means the user could be interested in that part of keywords, which is countable to clarify the interests of the user. The SIGQP then tries to find a k-user set $$U_{k}$$ , $$k \le n$$ , such that the union of keywords of these k users will equal to W, and additionally, the diameter (longest Euclidean distance of two arbitrary users in $$U_k$$ ) should be as small as possible. The SIGQP has been proved as NP-Complete, and two heuristic algorithms have been proposed. Extended from SIGQP, the main problem of this paper prioritizes in finding the smallest k for $$U_{k}$$ to cover all the keywords, with the users’ distance as the secondary criterion, called as “minimum user spatial-aware interest group query problem” (MUSIGQP). This paper further designs an exact algorithm on a measure-&-conquer-based method to precisely solve this problem, and a performance analysis is given.

In the time series classification (TSC) problem, the calculation of the distance of two time series is the kernel issue. One of the famous methods for the distance calculation is the dynamic time warping (DTW) with $$O(n^2)$$ time complexity, based on the dynamic programming. It takes very long time when the data size is large. In order to overcome the time consuming problem, the dynamic time warping with window (DTWW) combines the warping window into DTW calculation. This method reduces the computation time by restricting the number of possible solutions, so the answer of DTWW may not be the optimal solution. In this paper, we propose the minimum-first DTW method (MDTW) that expands the possible solutions in the minimum first order. Our method not only reduces the required computation time, but also gets the optimal answer.

Metric 1-median asks for $${\mathrm {argmin}}_{p=1}^n\, \sum _{q=1}^n\,d(p,q)$$ , breaking ties arbitrarily, given a metric space $$(\{1,2,\ldots ,n\},d)$$ . Let A be any deterministic algorithm for metric 1-median making each point in $$\{1,2,\ldots ,n\}$$ involve in only O(1) queries to d. We show A to not be $$o(\log n)$$ -approximate.

A (k, n) threshold secret sharing scheme encrypts a secret s into n parts (called shares), which are distributed to n participants, such that any k participants can recover s using their shares, any group of less than k ones cannot. A robust threshold sharing scheme provides not only the perfect security, but also the tolerance of a possible loss of up to n–k shares. When the size of s grows large (such as multimedia data), the efficiency of the encoding/decoding on s becomes a major problem. We designed efficient implementations for Kurihara et al.’s threshold secret sharing scheme on parallel GPU platforms in a personal computer. Experimental results show that the parallel GPU implementation could achieve an appealing speedup over the sequential CPU implementation when dealing with the sharing of multimedia data.

We consider capacitated vertex cover with hard capacity (HCVC) on f-hypergraphs. In this problem, we are given a hypergraph $$G=(V,E)$$ with a maximum edge size f. Each (hyper)edge is associated with a demand and each vertex is associated with a weight (cost), a capacity, and an available multiplicity. The objective is to find a minimum-weight vertex multiset, or cover, such that the demands of the edges can be met by the capacities of the vertices and the multiplicity of each vertex does not exceed its available multiplicity.In this paper we present an O(f) bi-approximation for partial HCVC. As the demand served is at least the ratio of $$(1-\epsilon )$$ , we have an $$O(1/\epsilon )f$$ -approximation algorithm. This gives a parametric trade-off between the total demand to be covered and the cost of the resulting demand assignment.

This paper presents an adaptive learning rate of recursive sinusoidal function neural network (ALR-RSFNN) with Lyapunov stability for identification elders fall signal. The older human signal analysis has been a research topic in health care fields that algorithms are implemented in wearable device real time to detect fall situation. However, the code size of the microcontroller in wearable device is limited, and the neural network learning rate choice is important which influencs neural network convergence performance. The recursive sinusoidal function neural network uses sine wave modulation input function to reduce train times in traditional Gaussian function vertex and width. Moreover, we utilize adaptive learning rate to guarantee network stability. In the experimental results, the ALR-RSFNN identify human fall signal accurately and reliably. In addition, we use wearable device combined BLE (Bluetooth low energy) to feedback output response real time.

This paper presents a multi-recursive wavelet neural network (MRWNN) with proximity capacitive gesture recognition. Recently, the capacitive sensor technologies have been developed for proximity methods that sensing electronic varies around sensor detection point, but the user gesture signals are time-variant. The MRWNN have multi layers recursive weight to record last signal variation, and we utilize microcontroller with MRWNN to identify algorithms and implement proximity capacitive gesture recognition. Moreover, we show MRWNN weight convergence analysis of the MRWNN signal identifier. In the experimental results, we show MRWNN can recognize patterns of different gesture signal accurately and reliably. In addition, we use wearable device combined with BLE (Bluetooth Low Energy) feedback output response immediately.

A paired-dominating set of a graph G is a dominating set S of G such that the subgraph of G induced by S has a perfect matching. In [Paired domination in graphs, Networks, 32:199–206, 1998], Haynes and Slater introduced the concept of paired-domination and showed that the problem of determining minimum paired-dominating sets is NP-complete on general graphs. Ever since then many algorithmic results are studied on some important classes of graphs. In this paper, we extend the results by providing an $$O(n^2)$$ -time algorithm on distance-hereditary graphs.

There are many kinds of edge colorings. This paper deals with a special edge coloring named rainbow coloring. Different from other edge colorings, the rainbow coloring requires every pair of vertices has a rainbow path between them. A rainbow path is a path $$ \, P \, $$ in graph $$ \, G \, $$ such that every edge on P has different color. A rainbow connected graph $$ \, G \, $$ is a graph such that there is a rainbow path between every pair of vertices. The rainbow connection number, denoted $$ rc(G), $$ is the smallest number of colors to meet the conditions that the graph $$ G $$ is rainbow connected. The bubble sort graph, denoted as $$ B_{n} , $$ is a type of Cayley graph. This paper established the rainbow connection number of bubble sort graph.

In this paper, we aim to provide a general study on the framework of Multi-Service Location Problems from a broader perspective and provide systematic methodologies for this category of problems to obtain approximate solutions. In this category of problems, we are to decide the location of a fixed number of facilities providing different types of services, so as to optimize certain distance measures of interest regarding how well the clients are served.Specifically, we are to provide p types of services by locating $$k \ge p$$ facilities. Each client has a demanding list for the p types of services, and evaluates its service quality by its service distance, defined as its total transportation cost to those facilities offering the demanded services. Under this framework, we address two kinds of distance measures, the maximum service distance and the average service distance of all clients, and define the p-service k-center problem and the p-service k-median problem, according to the minimax and the minisum criteria, respectively. We develop a general approach for multi-service location problems, and propose a (2p)-approximation and a 4-approximation to the two problems, respectively.

The total k-domatic partition problem is to partition the vertices of a graph into k pairwise disjoint total dominating sets. In this paper, we prove that the 4-domatic partition problem is NP-complete for planar graphs of bounded maximum degree. We use this NP-completeness result to show that the total 4-domatic partition problem is also NP-complete for planar graphs of bounded maximum degree. We also show that the total k-domatic partition problem is linear-time solvable for any bipartite distance-hereditary graph by showing how to compute a weak elimination ordering of the graph in linear time. The linear-time algorithm for computing a weak elimination ordering of a bipartite distance-hereditary graph can lead to improvement on the complexity of several graph problems or alternative solutions to the problems such as signed total domination, minus total domination, k-tuple total domination, and total $$\{k\}$$ -domination problems.

Given a metric graph $$ G = (V,E,w) $$ , a center $$ c \in V $$ , and an integer $$ p $$ , we discuss the Star $$ p $$ -Hub Routing Cost Problem in this paper. We want obtain a depth- $$ 2 $$ tree which has a root and root is adjacent to $$ p $$ vertices called hubs. We call that this tree is a star $$ p $$ -hub tree and let the sum of distance in tree between all pairs of vertices be minimum. We prove the Star $$ p $$ -Hub Routing Cost Problem is NP-hard by reducing the Exact Cover by $$ 3 $$ -Sets Problem to it. The Exact Cover by $$ 3 $$ -Sets Problem is a variation of set cover problem and known NP-hard problem. After proving the Star $$ p $$ -Hub Routing Cost Problem is NP-hard, we present a $$ 4 $$ -approximation algorithm running in polynomial time $$ O(n^{2} ) $$ for the Star $$ p $$ -Hub Routing Cost Problem.

Using Artificial Neural Networks, Random Forest and Support Vector Machines algorithms to optimize Tube inner circumference state classification and accomplish the process of Incoming Quality Control (IQC) is proposed in this paper. However, the traditional classification system is usually set the threshold by the developer in the early stages. The method is time-consuming and tedious to develop the module. In modern, machine learning technology can overcome the shortcomings of tradition classification system. However, machine learning exists a lot of algorithms, such as Artificial Neural Network (ANN), Random Forest (RF), Support Vector Machine (SVM), and so on. And, the different algorithms may cause the different characteristics and efficiencies, so it’s necessary to compare the different algorithms at application. This paper will use a method, called grid search to find the best parameter, and compare these algorithms which has the best characteristic, efficiency and the parameter. Finally, it is found from the experimental results that the method of this paper is workable for actual dataset.

In recently years, intelligent manufacturing system is getting grown up. There are many manufacturing vendors that investigate to develop their intelligent manufacturing systems in their product line. Our government also promotes some cities of mid Taiwan to become intelligent cities in the future. However, we discover that there are no such stand cryptography modules or security modules in these systems and there are a lot of papers that indicated security problems in intelligent manufacturing system. When a legimate user attempts to login an intelligent manufacturing system, an attacker may pretend as a legal server to perform mutual authentication with this user. On one hand, the server also faced the same problem. That is called the men-in-the-middle problem. On the other hand, an attacker may lunch software to capture some users’ login passwords. If a user’s login password does not to be protected by security mechanisms of the intelligent manufacturing system, the attacker may login this system with captured passwords successfully and fetch secret data in advance. Besides, there are no such provable user authentication schemes for the intelligent manufacturing system in random oracle model. Due to above problems, we proposed a secure user authentication scheme for intelligent manufacturing system. Not only this scheme does not store each user’s password in its database table, but also each user can change his/her own password after logging in successfully. Finally, we also give formal security proof in the random oracle model of the full version of this paper.

A strong designated verifier signature (SDVS) is a special variant of digital signatures, since it only allows a designated recipient to verify the signer’s signature. The transcript simulation property of such signatures also prohibits a designated verifier from arbitrarily transferring his/her conviction to any third party. When implemented in certificateless cryptosystems, a certificateless SDVS is unnecessary to manage public key certificates and deal with the key-escrow problem of conventional identity-based systems. In 2014, Shim pointed out a crucial security property called non-delegatability for SDVS schemes. This property states that anyone should not be able to generate a valid SDVS without obtaining either the signer’s or the verifier’s private key. In other worlds, a non-delegatable SDVS scheme must ensure that any malicious adversary cannot forge a valid signature even if he/she has gotten the shared secret value between a signer and an intended verifier. In this paper, we first demonstrate that a previously proposed efficient certificateless SDVS scheme is vulnerable to the delegatability attack and then further propose an improved variant.

While IoT becomes more and more popular, security becomes an important issue when IoT deployment. Considering there are lot of mobile device, it is frequent for member joining and leaving. Therefore, traditional key agreement schemes are not suitable for dynamic IoT environments. In this paper, we propose a dynamic key management scheme to avoid key update overhead when membership changing.

Not only does Named Data Network (NDN) provide a higher performance than TCP/IP network, but also it can cope with the problem of limited IP addresses. Currently, some scholars have proposed a secure file transfer protocol, which is based on re-encryption for NDN, to ensure that files can be safely transferred between nodes or users. On the other hand, because of the popular craze of saving files on Cloud, the security of Cloud is in vogue. However, numerous of jobs must be done if the user wants to calculate encrypted files on Cloud. Furthermore, the process will make files exposed in danger for a long time. By using Homomorphic Encryption, we can perform operations of the encrypted files on Cloud, and the operations will be finished by Cloud. In this paper, we provide secure operations for a file transfer protocol based on homomorphic re-encryption for named data networks. Our scheme fuses the advantages of NDN, Homomorphic Encryption, and Re-Encryption. It is a low energy consumption system which offers transmission safety and handy calculation.

Over the past years, a visual secret sharing scheme gave each share the same ability to reconstruct the secret image. That means each participant had the same significance to reveal the secret. However, it is not realistic in real world for some applications. In this paper, the weighted $$ (k,n) $$ -threshold visual cryptography scheme is proposed, in which the secret image is encoded into some shares with different weights and the sum of weights for all shares has to be equal to $$ n $$ , according to the share-holders of different priority levels. In the decoding phase, when the shares with different weights are superimposed and the sum of weights of all stacking shares is equal to or larger than $$ k $$ , the secret is then visually recognizable. Otherwise, no information about secret image is revealed. The proposed scheme is the first weighted $$ (k,n) $$ -threshold visual cryptography method with no pixel expansion. It is suitable for applications in the real life.

Recently, Fan et al. proposed the File Transfer Protocol Based on Re-Encryption for Named Data Network (FTP-NDN) in order to reduce the cost that affects simultaneous access of same video services. The authors designed an elegant network architecture to deal with secure file transmission to the unknown potential customers. The technique is shown to be secured under Decisional Bilinear Diffie-Hellman (DBDH) assumption and computationally efficient than other existing techniques. Although, the protocol achieves data confidentiality, it does not provide node authentication during transmission in the NDN. In this paper, we propose an authentication scheme using the bilinear pairing. Performance evaluation shows that the proposed technique can be incorporated with considerable computation overhead.

Everyone now can store their files to the cloud, which makes life more convenience and don’t worry about losing the storage device. However, the files store in the cloud is managed by someone may not trustable become a security concern. One of the solutions is encrypting the file and doing assured deletion on it. This paper proposed a time-based assured deletion in a flexible hierarchy structure. Clients will be distributed one derivation key depend on which class and when to over. Then client can use the derivation key to derivate the time-bound secret key to encrypt files. While the predetermined time passing, the secret key will be deleted and the file is unrecoverable. The proposed scheme provides the guarantee of files in the cloud, and also applies to the flexible hierarchy structure which may suitable in the organization.

With the widespread use of smartphones, many malware attacks such as user’s private information is stolen or leaking have been proposed. Furthermore, the hacker can manipulate these smartphones to become a member of malicious attackers. Therefore, how to detect the malware application has become one of the most important issues. Until now, two detection methods (static analysis and dynamic analysis) were discussed. For the static analysis view, it observes the source code to determine whether it is a malware application. However, the source code will be processed (such as packing or confusion) before it is shared. Therefore, the static analysis method is not able to detect it because we cannot get the recover code correctly and completely. In order to overcome this disadvantage, a new detection method based on CNN (convolutional neural network) will be proposed in this paper. The major contribution of our proposed scheme is that we can decompress the APK (Android application package) file directly, to obtain the classes.dex file and then uses the training detection model to determine whether the input classes.dex is malicious code or not. Finally, according to the experiment results, our proposed scheme is available for all APKs with an accuracy rate is 94%.

As the emerging threats of cybercriminals in recent years, how to efficiently and economically identify stealthy activities and attacks to avoid sensitive information leakage has been an important issue. However, due to business confidentiality and a lack of trust among information sharing, such valuable information is not exchanged transparently and not well utilized so far. In this study, we propose a hybrid method for internal threat identification. Our method leverages external open-source intelligence and applies it to internal network activities to uncover potential hacking campaigns among the network. We present the method consisting of collecting external intelligence, detecting internal infections, and identifying threats. We conduct our experiment under a tier-1 network in Taiwan. From the results, our method successfully identifies a number of famous hacking groups which are underneath threats in the large-scale network.

Early glaucoma diagnosis prevents permanent structural optic nerve damage and consequent irreversible vision impairment. Longitudinal studies have described both baseline structural and functional factors that predict the development of glaucomatous change in ocular hypertensive and glaucoma suspects. Although there is neither a gold standard for disease diagnosis nor progression, photographic assessment of the optic nerve head remains a mainstay in the diagnosis and management of glaucoma suspects and glaucoma patients. We describe a method aimed at both detecting pathologic changes, characteristic of glaucomatous optic neuropathy in optic disc images, and classification of images into categories glaucomatous/suspect or normal optic discs. Three different deep-learning algorithms used are transfer learning, deep convolutional neural network, and deep multilayer neural network that extract features automatically based on clinically relevant optic-disc features. Of the total of 455 cases extracted from the RIM-ONE public dataset (version 2), consisting of 348 training, 87 validation and 20 test cases, the proposed approach classified images with a training accuracy of 98.16%. We hypothesise that this approach can support the clinical decision algorithm in the diagnosis of glaucomatous optic neuropathy.

Many adjectives have been used to describe voice characteristics, yet it is challenging to define sound style precisely using quantitative measure. In this paper, we attempt to tackle the voice style classification problem based on techniques designed for speaker recognition. Specifically, we employ i-vector, a widely adopted feature in speaker identification, and support vector machine (SVM), for style classification. In order to verify the reliability of i-vector, we conduct pilot study, including noise sensitivity, minimum voice duration, and mimicry style test. In this study, we define eight voice styles and collect appropriate voice data to process and verify our hypothesis through the experiment. The results indicate that i-vector can indeed be utilized to classify voice styles that are commonly perceived in daily life.

Calligraphy is one of the most important cultural art as well as writing tool in ancient China. Various writing styles have evolved over time in Calligraphy text, including Regular script, Clerical script, Semi-cursive script, Cursive script, and Seal script. In this study, we consider the cursive Chinese calligraphy recognition task as a variant of handwritten text recognition. We apply deep convolutional network approach to this recognition problem and achieve 84.6%, 92.6%, 93.7%, 96.7% average top1, top3, top5, top10 accuracy for 395 characters and 83.8%, 91.8%, 94%, 96.1% average top1, top3, top5, top10 accuracy for 632 characters. Our investigation indicates that text recognition tasks can be tackled by deep learning based approach even only when a limited number of training samples are available.

In this paper, we streamline the Grassmann multivariate time sequence (MTS) clustering for state-space dynamical modelling into three umbrella approaches: (i) Intrinsic approach where clustering is entirely constrained within the manifold, (ii) Extrinsic approach where Grassmann manifold is flattened via local diffeomorphisms or embedded into Reproducing Kernel Hilbert Spaces via Grassmann kernels, (iii) Semi-intrinsic approach where clustering algorithm is performed on Grassmann manifolds via Karcher mean. Consequently, 11 Grassmann clustering algorithms are derived and demonstrated through a comprehensive comparative study on human motion gesture derived MTS data.

Inflammatory cells such as lymphocytes and neutrophils are crucial indicators in diagnosing acute inflammation from liver histology images. However, there are several challenges in detecting the inflammatory cells. The inflammatory cells have large variation and also appear similar to other cells. In an often occasion, the inflammatory cells may overlap each other. It is also unavoidable to see the clustery noise in the background. To conquer the above-mentioned problems, this paper proposes a procedure, which implements the detection-then-classification by combining the distance transformation with deep convolutional neural networks for detecting an accurate position of each cell. Then a precise image patch can be extracted for a deep convolutional neural network for classification of the cells into nuclei, lymphocyte, neutrophils and impurity (e.g. Kupffer cell). The experimental results show that the proposed approach can effectively detect the inflammatory cells from H&E Staining liver histopathological images, with an accuracy of 93.7% in inflammatory cells classification.

Now Internet of Things is growing fast and presents huge opportunities for the industry, the users, and the hackers. IoT service providers may face challenges from IoT devices which are developed with software and hardware originally designed for mobile computing and traditional computer environments. Thus the first line of security defense of IoT service providers is identification of IoT devices and try to analyze their behaviors before allowing them to use the service. In this work, we propose to use machine learning techniques to identify the IoT devices. We also report experiment to explain the performance and potential of our techniques.

In this paper we propose a deep learning based depth estimation method for monocular RGB sequences. We train a pair of encoder-decoder network to resolve depth information form image pairs and relative camera poses. To solve scale ambiguous of monocular sequences, a conditional generative adversarial network is applied. Experimental results show that the proposed method can overcome the problem of scale ambiguous and therefore is more suitable for a variety of applications.

Trajectory computation for forklifts and pedestrians is of relevance for warehousing applications such as pedestrian safety and process optimization. We recorded a novel dataset with a varying range of forklift models and pedestrians, busy with loading or unloading in warehouses. We have videos with frequently occluded trucks in aisles and besides racks, some with busy pedestrian activity, such as in docking areas. Robust target localisation is very essential for seamless tracking results. For localising forklift trucks/pedestrians, we trained a deep-learning based, faster region-based convolution neural network (faster RCNN) on our own recorded data. We used detection from the model output to configure a Kalman filter to estimate the trajectories in the image plane. We also improved the forklift trajectory based on computing pixel saliency maps for the region of interest detected by faster RCNN. Our analysis shows that with robust target detection (fewer false positives and false negatives) from our trained network and Kalman-filter-based state correction, tracking results are close to ground truth.

The air pollution has become a major ecological issue. The surpassed pollution levels can be controlled by searching the pollution source. An environmental monitoring unmanned aerial vehicles (UAVs) can address this issue. The challenge here is how UAVs collaboratively navigate towards pollution source under realistic pollution distribution. In this paper, we proposed a novel methodology by using the collaborative intelligence learned from Golden shiners schooling fish. We adopted shiners collective intelligence with the particle swarm optimization (PSO). We used a Gaussian plume model for depicting the pollution distribution. Furthermore, our proposed method incorporates path planning and collision-avoidance for UAV group navigation. For path planning, we simulated obstacle rich 3D environment. The proposed methodology generates collision-free paths successfully. For group navigation of UAVs, the simulated environment includes a Gaussian plume model which considers several atmospheric constraints like temperature, wind speed, etc. The UAVs can successfully reach the pollution source with accuracy using the proposed methodology. Moreover, we can construct the unknown distribution by plotting the sensed pollution values by UAVs.

In the current trend, design pattern has been widely used for improving software quality. However, using design patterns is not easy, developers need to understand their complex structure and behavior, and have to apply them in the correctly. Therefore, several approaches are proposed to check the violence of pattern application in a system. We argue only static checking to patter structure is not enough, dynamic testing to test the patterns’ semantics is necessary. In this paper, we propose a test model for design patterns. We explore the potential error point of each design pattern thoroughly, and encapsulate their testing methods into a package called “test pattern for design pattern (TP4DP)”. Just like the design pattern, our proposed TP4DP includes the sample code of executable test cases which facilitate their application practically.

Test case generation for arrays is more sophisticated than scalars. It involves the generation of both the size of an array and the values of the array elements. This issue is more challenging in black-box testing than in white-box testing because the specification usually does not describe how arrays are processed in the program. This paper proposes a constraint-based approach to generate test cases for Java arrays in black-box method-level unit testing. The constraint-based framework in this paper uses Object Constraint Language as the specification language. The constraint-based specification is then converted into a constraint-based test model, called constraint logic graph. A constraint logic graph is a succinct representation of the disjunctive normal form of the specification. Test case generation is formulated as a set of constraint satisfaction problems generated from the constraint logic graph. These constraint satisfaction problems are then solved using the constraint logic programming to generate the test cases.

With the evolution of software engineering technology, using cloud services to replace self-built information systems has been proven an economical and reliable way. However, how to help e-commerce service system builders to choose suitable compositions of cloud services that meet their needs is still a challenge. In the past decade, a number of academic studies have explored the selection strategies and algorithms for cloud services, however, most efforts are not able to consider multiple types of cloud services simultaneously to provide composite cloud service solutions. To address the above issue, this study proposes a cost-driven recommendation method, called ECSSR (E-Commerce Service Suite Recommendation). ECSSR takes the user budget as the core factor and simultaneously considers the user’s preferences for the service types. A prototype system, referred to as ECClouder, is also designed and implemented to realize the features of ECSSR. ECClouder is able to collect the user’s requirements, convert application-level requirements into infrastructure-level requirements, and produce appropriate cloud service solutions. The case study show that ECClouder can effectively help users to find cloud service solutions that are reasonably priced and meet their needs.

Breast cancer has become the third leading cause of death for women in Taiwan. For clinical pathologists, the grading criteria: Nottingham Modification of the Bloom-Richardson (NBR) System based on histological pathology is a gold standard to assess the lesion severity of the invasive ductal carcinoma. The grading indices for the disease based on NBR include tubular formation, pleomorphism, and mitotic count. Because the manual grading is measured depending on qualitative analysis, it usually causes a big workload due to its various variability. The major goal of this work is to extend our previous work and propose a computer-aided-diagnosis system to assess quantitatively the severity of the breast carcinoma. To this end, it first analyzes the H&E stained slide images of the breast specimen using a series of image processing operations to extract feature parameters related to morphometry of mammary tissue, and hyperplasia degrees of nucleus, and mitotic count of nuclei based on histology and cytology, and choosing important features with feature selection, and identify the scores using support vector machine finally. Experimental results reveal that the proposed system not only can obtain satisfactory performance, but also provide histological grade and prognosis information for clinical pathologists to improve the efficiency of diagnosis.

Objective: We propose Cateye, a Python-based search engine framework tailored for searching in biomedical classification systems such as ICD-10, DSM-5, MeSH, and SNOMED CT. Many of the biomedical classification systems have coarse-grained and fine-grained structures to handle the different levels of information. The general-purpose search engines, which designed for document retrieval, face three major problems: too strict terminology, not efficient search, and uncertainty to stop searching. These disadvantages make it painful to search in the classification systems.Materials and Methods: We used the ICD-10 coding systems as our sample materials. We design a hint bar which shown along with the search results and dramatically helps the users to formulate the correct query. A hint is a suggestion of search term which can best divide the search space into half-and-half.Results: The case studies show that our hint mechanism performs at least one step deeper per search step in most cases.Conclusion: The source code of Cateye for searching the classification systems associated with coarse-grained and fine-grained architecture is available at https://github.com/jeroyang/cateye .

On the wearable smartglasses device, this paper proposes a simple but practical 2D ear detection algorithm based on Arc Mask Superposition Operator (AMSO) and luminance density verification. In detail, in the first half phase of the proposed ear detection algorithm, a few ear candidates are extracted by AMSO followed by multilayer mosaic enhancement and orthogonal projection histogram analysis. Then, in the second half phase, the most likely ear candidate can be effectively verified by a straightforward comparison of luminance density. Experimental results show that the proposed ear detection algorithm without any detection false positive can achieve better hit rate and faster response performance than conventional AdaBoost-based ear detection algorithm. Afterward, Coherent Point Drift feature extraction algorithm on Android smartglasses device is also introduced. Implementation results show the real-time performance of the wearable ear recognition smartglasses is feasible for diverse biometric applications.

Ultrasound imaging is the most commonly applied method for the diagnosis and surgery of a trigger finger. However, the ultrasound images are noisy and the boundaries of tissues are usually very unclear and fuzzy. Therefore, an automatic computer assisted tool for the tissues segmentation is desired and developed. The segmentation results of the conventional methods were satisfactory but they usually depended on the prior knowledge. Recently, the deep-learning convolutional neural network (CNN) shows amazing performance on image processing and it can process the image end-to-end. In this study, we propose a finger tendon segmentation CNN which overcomes the requirement of prior knowledge and gives promising results on ultrasound images. The evaluation result is remarkable high with DSC 0.884 on 380 testing images and the prediction time is fast by 0.027 s per image. This work, to our best of knowledge, is the first deep learning finger tendon segmentation method from transverse ultrasound images.

In the diagnosis of prostate disease, urologists examine the presence of fibrosis or tumors in the prostate tissue of patients by magnetic resonance imaging (MRI) and transrectal ultrasonography (TRUS). In general, urologists will imagine the location of fibrosis or tumors in 3D space. To resolve this a fusion algorithm for three-dimensional (3D) prostate magnetic resonance imaging (MRI) and 2D TRUS was developed. Fiducial registration error was employed in this study to evaluate the registration error of the upper half contour. Experiments showed that the proposed method achieved good fusion results. The system can further guide urologists in the performance of TRUS for tissue biopsy or high intensity focused ultrasound (HIFU).