Earticle

Masonry is a common type of construction that uses mortar to bind individual units, such as brick or building stones, together to construct the structure. Even though masonry structures are durable, multiple factors such as the quality of mortar, workmanship, and harsh environment could greatly reduce the structural integrity, leading to defects and even human loss. Thus, it is crucial to perform the maintenance process regularly. Previously, the maintenance relied mainly on inspectors, who inspected the masonry structures to find cracks and determine the seriousness. However, this process is error-prone, costly, and time-consuming. As a result, this study proposes a fully automated masonry crack segmentation framework that robustly identifies various types of masonry cracks. In addition, the length of the segmentation cracks, which has been ignored in previous studies, is also computed.

Often most of the modern human people are suffering from a long time of working or studying on the stationary pose. Subsequently, the health of our life is highly threatened to be exacerbated by chronic orthopedic diseases. In order to solve this social problem, we suggest pose detection that can have the people who have deleterious postures be notified. By using nowadays advanced computer vision techniques, in this paper we suggest the posture recognition module to enhance our quality of life. While most posture recognition recognizes only one person's posture, we made our pipeline to perform posture recognition for multiple people through images obtained through a single camera. One of the big problems in measuring people's postures is that it is necessary to distinguish the various body structures and postures of people. For this, posture images and labeling of various people are required. We created pose images of people of various body types through images of a small number of people through skeleton-based coordinates augmentation. We made a posture classifier using various models and observed the improvement of augmentation performance for each model. Through this, we found that the postures of various people can be measured using a relatively small data set. In particular, for deep-learning models that require a lot of data, generalization performance was greatly improved.

With the rapid development of social media, users' next location prediction has become an important research direction, which can provide personalized travel suggestions for users. However, existing methods ignore the semantic relationship between users' historical and current trajectories. This paper proposes a new method for predicting the user's next location to solve this problem. We first process the user POI data as trajectory data, use the attention mechanism to extract similar features of the user's historical trajectories, and then combine them with the current trajectory features to obtain the user's next location recommendation. The experimental results show that our proposed model performs satisfactorily on a real dataset.

Malware detection has piqued the interest both academia and anti-malware industry as a result of the malware explosive growth over the past 20 years and the havoc that it has been able to cause. Even though in the past signature-based anti-virus systems have been successful, malware authors and cyber security experts have since been in a never-ending arms race. In order to overcome the endeavors of cyber security experts, malware authors created polymorphic, metamorphic, and oligomorphic engines for malware in order to bypass the detection of traditional anti-virus systems. As a result, cyber security experts sought to devise their best strategies for retaliating against adversary. Conventional algorithms of machine learning and more complex ones of deep learning constitute the remedy to such impediment. The major vulnerability of machine learning-based malware detection systems is represented by adversarial examples. In this paper, we propose a machine learning-based malware detection system that is resistant to adversarial malware by utilising code normalisation. We evaluate the impact of code normalisation in a deep learning based-malware detection system and the proposed malware detection system with the code normalisation scored 99.02% success rate.

Recently, research on pedestrian action recognition from the vehicle’s viewpoint is being studied in many ways. The information about pedestrian action classification is very important for autonomous driving to determine safe path planning and avoid accidents. To provide a computationally efficient solution to pedestrian action recognition, this paper proposes a multi-head CNN model to currently extract multi-actions of pedestrians from the unified model. This model consists of one pre-trained backbone network and two head networks. One head network classifies Gait (walking/standing) and the second classifies Attention (looking/non-looking) of pedestrians. The proposed model offers a lighter model with smaller memory, faster processing speed, and alleviates data imbalance problem – a common problem found in most of dataset – leading to improved accuracy.

Multi-object tracking techniques are receiving increasing attention due to the growing demand of autonomous driving. Recently, the performance of multi-object tracking has been improved significantly thank to deep learning technique. Most of multi-object tracking methods based on deep learning, however, are highly prone to frequent tracking losses and track-ID switching in case of limited viewpoint and occluded objects. To alleviate this problem, we propose a multi-camera Collaborate Multi Object Tracking (CMOT) method which performs online association of multiple tracked vehicles from stereo vision camera. CMOT not only provides global tracking IDs between multiple cameras but also helps reduce the problem of ID switching compared with the conventional multi-object tracking based on single camera. It can, therefore, improve the overall performance of multi-vehicle tracking compared to each individual camera. We demonstrate the multi-object tracking performance of the proposed method using stereo images of the KITTI dataset.

A story generation task is to develop a system that can continuously generate natural, consistent, and coherent stories for consecutive scenes. Recently transformer-based language models have shown considerable results at the sentence-level generation for learning human-writing ability. However, it is very crucial to understand the way of developing the story using a combination of various contents. Recent works have mainly focused on human-guided AI story generation methods in which humans as guidance determine the next storyline, and the system creates a story that reflects the storyline well. This study focuses on the way of replacing the human role with the AI-based model. Based on this, this study deals with the methodology for creating a long story spanning multiple scenes rather than creating a story at the level of one scene. In this regard, we propose a novel AI-guided story generation framework with automatic storyline generator. It is a pipeline structure consisting of two modules such as a storyline generator and a story generator, which enables the continuous creation of coherent stories. Particularly, we transform the storyline generation problem into a multiple-choice QA problem to predict the next storyline. This study shows the possibility of generating continuous stories for multiple scenes without any human intervention.

Various voice disorders exist in the world, and many studies on voice pathology detection have been conducted. Voice pathology detection (VPD) has made various advances by medical examination, but it is necessary to apply artificial intelligence (AI) to VPD for quick and convenient diagnosis of suspected patients and efficient use of expert resources. Recently, research to detect it using artificial intelligence has been studied. In the research of biomedical engineering, automatic VPD system by machine learning algorithms and well-established features has become a research hotspot. We used the VOice ICar fEDerico II (VOICED) dataset, which has been widely used in the VPD system. It contains 150 pathological voices and 58 healthy voices, resulting in class imbalance. In this paper, we try to figure out the degree of accuracy improvement by using the oversampling technique and multiple models to automatically detect and classify pathological voices in the class imbalanced dataset.

Brain tumor is a very terrible disease. Brain tumor is caused by an increased number of cells. The presence of the skull layer around the brain makes it tough in studying the behavior of growth cells. It also raises the complication for the identification of disease. The initial discovery of a brain tumor is necessary to defend the survival of patients. Frequently, the brain cancer segmentation, and classification through the MRI images technique. Though, the radiologists are not providing actual visualization of brain cells in MRI images due to the irregular growth of cells, which forms of cells are growing rapidly and slow at some stage in brain tumors in the brain. So, automatic strategies are required to evaluate thoughts tumors exactly from MRI images in this research automatic, MRI brain tumors are used for classification, segmentation, and Behavior analysis of cell growth. The problem of visualization of cell growth and behavior analysis of brain cells is solved through MRI images which enhance the detection of cancer. To analyze the behavior of cell growth, which forms of cells are growing rapidly and slow at some stage in brain tumors, and analyze the area of images in which type of cells is affected. Single models are less efficient. We will use ensemble models which would also be helpful for better performance and accuracy.

Electric energy is the basic need for human survival on this earth as these needs increase with the rapid increase in population. It’s become a challenge to manage home energy with the current situation. Smart grid provided different techniques to overcome these challenges to meet the need. This paper presents the result of the different optimization techniques that give the best performance in reducing cost, PAR, and user discomfort. Based on results the best result techniques are also combined to make a hybrid model for more accuracy. This paper not only describes optimization techniques but also the limitations and features of these techniques.

To train deep learning models faster, distributed training on multiple GPUs is the very popular scheme in recent years. However, the communication bandwidth is still a major bottleneck of training performance. To improve overall training performance, recent works have proposed gradient sparsification methods that reduce the communication traffic significantly. Most of them require gradient sorting to select meaningful gradients such as Top-k gradient sparsification (Top-k SGD). However, Top-k SGD has a limit to increase the speed up overall training performance because gradient sorting is significantly inefficient on GPUs. In this paper, we conduct experiments that show the inefficiency of Top-k SGD and provide the insight of the low performance. Based on observations from our empirical analysis, we plan to yield a high performance gradient sparsification method as a future work.

Autonomous driving relies on an accurate perception system that provides knowledge about surroundings and ensures safe driving performance. Usually, the perception system takes input information from onboard sensors (camera, LIDAR, RADAR, etc.) and then uses it to perform object detection tasks to accurately determine objects such as pedestrians, vehicles, traffic signs, and road barriers located around the ego vehicle. In order to have a safe trip and maneuver on the road, a vehicle detection algorithm should constantly improve the accuracy of vehicle detection. Since most of the conventional deep learning methods for vehicle or object detection rely on offline training with human-labeled large datasets, the conventional training methods have serious limitations in developing a breakthrough technique for gradual improvement in the detection accuracy of deep learning models. Thus, we propose a self-supervised training (SST) scheme that can gradually enhance detection accuracy with pseudo labeling.

A safe and robust autonomous driving system relies on accurate perception of the environment for application-oriented scenarios. This paper proposes deployment of the three most crucial tasks (i.e., object detection, drivable area segmentation and lane detection tasks) on embedded system for self-driving operations. To achieve this research objective, multi-tasking network is utilized with a simple encoder-decoder architecture. Comprehensive and extensive comparisons for two models based on different backbone networks are performed. All training experiments are performed on server while Nvidia Jetson Xavier NX is chosen as deployment device.

As the number of vehicle drivers is increasing day by day, the risk of traffic accidents is also increasing. Among the accidents, there are tire-related accidents causes huge damage if it occurs. These kinds of accidents could be prevented through safety checks of tire, but drivers usually overlook it because they don’t have the knowledge to know what the condition of the tire and don’t want to spend time and money to safety inspection and so on. To solve these problems, we propose tire life prediction mobile application with deep-learning method to check the condition of tires simply. Also, considering the embedded environment that has low power and capacity, we apply lightweight technique called pruning

The explosive growth of embedded vision enables the use of imaging products such as cameras and displays in IoT and multimedia applications. Mobile Industry Processor Interface (MIPI) Camera Serial Interface 2 (CSI-2) is the most commonly used interface to connect image sensors or displays to application processors or SoCs in such applications. To ensure that imaging modules such as cameras work correctly when integrated into a system, verification processes such as bus functional verification and bus performance verification are essential. This paper introduces a bus performance verification approach for MIPI CSI-2. By evaluating the performance of a two-camera system with different video parameter sets, the best possible performance of the system is obtained. Simulation results show that the bandwidth utilization of each camera is more than 99.5%.

Autonomous driving algorithm aims to safely control the vehicle through recognition of the surrounding environment and relevant decision and control based on deep learning algorithms. However, the deep learning algorithms require appropriate datasets but many difficulties exist in generating proper datasets for critical scenarios, e.g. accident avoidance. This paper presents a method for generating datasets for critical scenarios via a game engine. The video game engine used in this paper offers an open world environment, providing realistic graphics. A variety of non-player characters in the game equipped with a high level of Artificial Intelligence (AI) are ideally suited for the reproduction of various unexpected situations in reality by the interaction between the AIs. Using this, this paper creates a high-level critical scenario dataset and presents an augmentation method for the generated dataset.

It has been suggested that the main battery voltage of electric vehicles (EVs) is increased from 400 to 800 V to improve the charging speed and driving distance. However, the secondary components in the dc/dc converter of the on-board charger (OBC) have high voltage stresses due to the increased battery voltage. Therefore, this paper proposes a double-voltage charging technique. The previously designed and optimized dc/dc converter for 400 V battery is used to charge 800 V battery. The battery stack is divided into two modules. Dc/dc converter charges one of two modules alternatively using the battery selection circuit (BSC). Since the BSC has low voltage stresses, the proposed converter can have low voltage stresses on all the dc/dc and BSC parts. Moreover, the dc/dc converter and BSC can be decoupled by turning off the dc/dc converter while changing the charged battery. Therefore, the loss of BSC can be minimized by operating it with low frequency because the switching frequency of BSC can be independent of that of the dc/dc converter. The design process can also be simpler owing to decoupling operation. The LLC converter was adopted for dc/dc converter with 3.3 kW output to verify the proposed technique.

This study presents model-free reinforcement learn ing methods for economic and ecological adaptive cruise control (Eco-ACC) of connected and autonomous electric vehicles. For model-free optimal control of Eco-ACC, we applied two reinforcement learning methods, Deep Q-Network (DQN) and Deep Deterministic Policy Gradient (DDPG), in which deep neural networks of actors and critics were trained using IPG CarMaker simulations. For performance demonstrations, the HWFET, US06, and WLTP Class 3b driving cycles were used to simulate the front vehicle, and the energy consumptions of the host vehicle and front vehicle were compared. In high-fidelity IPG CarMaker simulations, the proposed reinforcement learning- based Eco-ACC methods demonstrated approximately 3–5% and 10–14% efficiency improvements in highway and city-highway driving scenarios, respectively, when compared with the front vehicle. A video of the CarMaker simulation is available at https://youtu.be/DIXzJxMVig8.

In this study, we consider application of deep learning methods in the cryptanalysis of tiny DES algorithm, which is a DES-like cipher. We develop two types of deep learning architectures to perform the cryptanalysis of tiny DES. It is a known-plaintext attack where the deep learning models only need ciphertext and plaintext pair as training and the learning target is to predict correct plaintext when a ciphertext is given. Simulation results have shown that deep learning methods cannot 100% recover the plaintext of tiny DES but can greatly reduce the analysis difficulty for plaintext recovery.

Nowadays, energy management and its optimization using smart devices are getting more attention due to their significant applications. Moreover, the applications used in these devices play a key role in developing smart cities that is only the way to solve urban problems. The potential of renewable energy sources like solar and wind power has been integrated in the smart grids to overcome the lack of supply via conventional fossil fuels and their environmental disputes that reduce operational cost. This review paper describes the significance of renewable power data that directly assists all the functions in smart cities such as the evolution of microgrids, renewable resources, energy forecasting, and power storage technologies. Furthermore, solar and wind power plants’ data with weather information as an additional cue is collected from different companies in South Korea. We aim to assist the researchers to develop artificial intelligence (AI)-based algorithms for power forecasting and establish its efficient management between suppliers and consumers.

Colonoscopy is the most effective examination way to detect colon polyps, which are highly related to colorectal cancer. Consequently, it is an important step to segment the poly accurately for diagnosis in clinical practice. However, most prior works focus on performance improvement using deep convolutional neural networks while the discrepancy between the training dataset and the test dataset is ignored. These distribution discrepancies may lead to the model overfitting the training dataset and lacking generalizability on unseen target domains. To alleviate this issue, we propose a Randomized Local Illumination Enhancement Network for polyp image segmentation. Specifically, we first employ an illumination decomposition network to decompose the input images into an illumination component and a reflectance component. The illumination component is augmented by randomly selected local illumination. Then the randomized local illumination-enhanced images are obtained by combining the augmented illumination and the reflectance, which are fed as the input of the segmentation network for improving the model generalizability. We conduct both quantitative and qualitative experiments on four polyp segmentation datasets. The satisfying results demonstrate the effectiveness of our proposed approach in the improvement of model generalizability on unseen data.

The recognition of crop diseases and pests based on images is one of the required techniques to identify damage due to diseases and pests and to take efficient management and prior actions. With the development of deep learning technology, the image recognition of diseases and pests using deep learning exhibited excellent performance and plays an important role in managing and controlling diseases and pests of crops. However, research on image recognition of diseases and pests in crops is facing difficulties due to the lack of large-scale datasets about diseases and pests. To solve this problem, this study proposes a system to manage and annotate the images of diseases and pests that can efficiently manage collected disease and pest images to generate high-quality disease and pest datasets. The proposed disease and pest image management and annotation system can collect disease and pest images uploaded from various sources and create statistics of them. It can also provide image inspection and user-friendly annotation functions.

In this paper, we propose a new framework that enables an object detector trained with only point-level annotations to estimate the centroids and sizes of objects in dense scenes. Specifically, the framework is based on the Swin Transformer structure and introduces a self-designed resolution feature fusion module in the hierarchical structure, where the estimation of object centroids is done directly by point supervision, and the object pseudo-size is initialized based on the assumption of local uniform distribution, and the regression of object size is guided by an improved congestion-aware loss function. In the NWPU-Crowd dataset, our method outperformed the existing state-of-the-art detection counting methods in F1-measure, precision, MSE evaluation criteria.

Rapidly growing innovative technologies enabled human beings to enjoy smart city services despite the development of such cities are still facing several challenges needed to be addressed. The waste management in smart cities particularly its segregation by smart methods is one of the primary concerns as the amount of waste generated every day by citizens is increasing. A comprehensive intelligent waste management system is direly needed to address the situation. This article aims to segregate recyclable and non-recyclable types of garbage collected from smart cities using the Intelligent Agent proposed and developed so far. The expected smart solution should provide the best level of accuracy at the lowest possible cost. Our study proposed a model to differentiate and segregate waste into recyclable and organic objects based on Intelligent Agent developed using a Convolutional Neural Network (CNN). The model proposed comprises of Intelligent Agent developed and the existing CNN model which is commonly used for transfer learning. The classification accuracy achieved is up to 93.27% which is better than the already published results of different models discussed in the recent past research studies. Furthermore, how can recyclable and organic waste be utilized in the future is part of our ongoing study. The findings may be of interest to practitioners and the researchers’ community working in the relevant field.

The spatial orientation of fire ignition can support an automatic fire suppression system (AFSS). However, due to the ambiguity of 2D images, most of the existing methods tend to study fire detection and cannot carry out spatial orientation. To solve this problem, we propose a spatial orientation method of indoor fire ignition based on monocular vision and virtual scenes. First, a hierarchical virtual scene construction method is proposed to realize the rapid construction of indoor scenes. Second, the characteristics of the fire are analyzed to obtain fire ignition in a 2D image. Finally, the spatial orientation of indoor fire ignition is calculated by analyzing the characteristics of fire attachment and the principle of three-dimensional imaging. The experimental results show that the absolute error of the spatial orientation of our method is 10.27 cm, and the relative error is 4.08%, which can meet the needs of AFSS.

Nowadays, renewable energy resources such as Photovoltaic (PV) is one of the convenient ways to integrate it into the distributed grid to fulfill the huge energy demands without burning costly and pollutant fossil fuels. Researchers have been contributing from various aspects to develop accurate PV-power forecasting methods however further improvements are needed for an effective power management system. Therefore, in this work, we propose an attention-based deep learning (DL) model (PV-ANet) for short-term PV-power forecasting. The proposed system mainly consists of three modules. First, data from an actual PV power plant is acquired and preprocessed to remove outliers and normalized for efficient processing. Next, the PV-ANet model is developed, which is consisting of an encoder and decoder modules. The encoder encodes the input attributes via stack conventional and attention layer. While the decoder part contains the normalization and series of the dense layers to expends the encoded features into optimal features and generate one hour ahead forecast. Finally, the proposed model is evaluated via standard error metrics including MSE, MAE, and RMSE and achieved the lowest errors rates compared to state-of-the-art methods.

Within cyber-physical networks, the Internet of Things is a futuristic idea, rich in promise as well as multifaceted requirements and implementation issues. Agent-Based Computing represents suitable and effective modeling, programming, and simulation paradigm for properly addressing them and fully supporting IoT system creation. Agent metaphors, principles, strategies, processes, and tools have all been used extensively in the development of IoT systems. In this paper, we have presented surveys and reports on the most recent contributions in this field.

As a massive number of real-time news makes it difficult for users to find their preferred news, various news recommender systems have been actively proposed in the research field. With the two popular real-world datasets in a news domain, Adressa and MIND, we compare the four state-of-the-art news recommendation methods (i.e., NRMS, LSTUR, NAML, and CNE-SUE) in terms of accuracy. Also, we investigate the strengths and weaknesses of news recommendation methods depending on datasets or metrics.

With the development of drone-related technology, drones are being used in various fields. Drones with resource-constrained characteristics have a problem with conducting a mission. The recovery process needs a lot of effort and cost. In this paper, we proposed a lightweight checkpoint system to reduce the recovery time and cost. The proposed energy consumption model for checkpoint-based drone recovery systems calculates the energy consumption of the drone according to the checkpoint period. We also confirm that there is an optimal checkpoint period to minimize the energy consumption of the drone.

The individual markets for indoor localization service (ILS) and augmented reality (AR) are expected to grow significantly. These combined services will be very diverse, and the market is expected to grow significantly. This paper presents representative technologies supporting ILS based on wireless communications and computer visions. Then, AR-enabled ILS cases are also presented.