Earticle

Scene Text Detection and Recognition is a problem of reading letters from everyday landscape images. Recently, STR research using deep learning models has been actively conducted for various applications. However, STR still has many challenges to overcome due to the complexity in image contents, which is much more challenging compared to recognizing letters in paper documents. In addition, the STR performance appears good for foreign characters such as English or Chinese, but poor for Korean. In this paper, we develop a model suitable for the Korean environment through the real world Korean scene text dataset. We measure the performance of various configurations of networks to find suitable network structure.

Recently, hand gesture recognition based on deep 3D convolution neural networks has made great progress. However, the large number of weight parameters that need to be optimized leads to its expensive computational cost. We introduced a transformer-based framework for hand gesture recognition, which is a fully self-attentional architecture. The framework abandons the conventional methods that rely on 3D convolution and proposed an approach to classify actions by focusing on the entire video sequence. In addition, we use a lightweight hand detector to continuously sample the video only when a hand is detected in the video sequence, thus reducing the computational consumption of the system. Experiments on two human hand gesture recognition benchmark datasets show the superiority of the proposed method, compared with existing state-of-the-art methods.

Automated speech emotion recognition (SER) by efficient long-term temporal context modeling is a challenging task of the digital audio signal processing domain. However, by default, the recurrent neural network (RNN) is employed to incorporate the temporal dependencies in sequence to investigate the relationships among sequences and features. In this study, we design a parallel convolutional neural network (PCNN) for SER by using a squeeze and excitation network (SEnet) with the self-attention module. Additionally, we adopt the residual learning strategy in both module, SEnet and self-attention, which is further improve the performance of the network. Our proposed SER system utilizes speech spectrogram as input and extracts utterancelevel discrete features by using the PCNN model. We experimentally evaluated our proposed system by standard speech corpus, interactive emotional dyadic motion capture (IEMOCAP). The prediction result reveals the significance and robustness of the proposed PCNN system, which obtained a high recognition rate of 72.01% over state-of-the-art (SOTA) methods.

In the agricultural field, the application of deep convolutional neural networks is increasing. Especially, in the task such as harvesting, instance-level segmentation is required to target fruits. Even though a large amount of data is required to train instance segmentation, it is not easy to obtain sufficient dataset for tomatoes. Therefore, synthetic images are generated through data augmentation for tomato instance segmentation. The training is performed through small number of real images and augmented images. As a result of training from real images, the best accuracy is 73.47%. Based on the synthetic data augmentation, the best accuracy is 89.87% with the generation of maximum 3 foreground objects per an image. We also show that the results of tomato detection and instance segmentation qualitatively.

In this paper, we present a new approach to time series forecasting, especially predicting stock data. Despite there have been lots of attempts for predicting future stock prices by using machine learning, performances have not been fine since many noises exist in stock data. In this work, we develop a novel method that using auxiliary classifier generative adversarial network to predict future stock prices. Basically, generative adversarial network suffers noise as an input of generator. This means generative adversarial network can be trained efficiently with the noisy data. In practice, our new method shows remarkable results compared to conventional other methods.

In the classifier which is constructed with a fully connected layer connected to an output layer the weight vector W creates a decision boundary between each. The learning of a feature vector influences the model to extract unique features of the class during the learning process which moves the distribution of the initial feature vector into the decision boundary of the class. When the similarity between the initial feature vector and the weight vector is high, it can be expected that the effect on the model is low because the loss value is small. On the other hand, if the similarity is low, it means that the interval between the feature vector and the weight vector is large and the loss value is high, which can be expected to have a higher effect on the model than when the similarity is high in the learning process of the model. In this paper, we verify how much the similarity between the initial feature vector and the weight vector before learning affects model learning. In order to confirm the effect of similarity, the model is learned by assigning an arbitrary fixed value so that the weight vectors W of the fully connected layer make different similarities. Both VGG16 and VGG19 models are used to compare the Recall and the Precision values of the class as a learning result of each model.

AI technologies are being applied in many areas as well as in modern malware detection technologies. However, adversarial attacks to such AI-applied malware detection technologies become one of major problems as it is in computer vision with AI. There has been a lot of effort on developing adversary-robust malware detection technologies, but it remains immature yet. In this paper, we review the state-of-theart in AI-applied malware detection technologies to identify limitations and shortcomings. We also present some directions for future research enhancing robustness against adversarial attacks.

The Internet of Things has become a new sensing paradigm for interacting with the physical world. As the sensors in the Internet of Things are often deployed in harsh environments, this makes the sensors prone to failure and malfunctions, producing abnormal and erroneous data, known as outliers. Anomaly detection is critical in the Internet of Things to ensure the quality of data collected by sensors by detecting a high probability of incorrect reads or data corruption. Because the energy of sensor nodes in wireless sensor networks is limited, the transmission between nodes in centralized anomaly detection will consume a lot of energy. Therefore, we propose a distributed anomaly detection method based on a mixture of low precision and high precision to save node energy and improve network life.

Watermarking techniques are mandatory to secure digital communication. For optimal technique, a high signal to noise ratio and normalized correctional is required. In this paper, a method that belongs to digital watermarking is proposed based on chaotic map through image gradient and least significant bit. The image is segmented into un-correlated blocks and calculate the gradient of each block. The gradient of the image expresses the rapid changes in an image. When watermark used chaotic substitution box (S-Box), it scrambled based on piecewise linear chaotic map (PWLCM). PWLCM has positive Lyapunov exponent and better balance property as compare to other chaotic map. This suggested s-box technique is capable to produce dispersive sequence with high nonlinearity in the generated sequence. By modifying the least significant bits of the original image the watermark signal is embedded based on the image gradient. Image block gradient ,direction and magnitude decide how much embeding can be done. The embedding payload compromises between robustness and imperceptibility. By the reverse operation, a watermarked image is obtained. In comparison with other methods, experimental results show satisfactory progress in robustness against several image processing and geometrical attacks while maintaining the imperceptibility of the watermark signal.

The ionosphere is the part of the Earth's atmosphere with a high concentration of free electrons and ions. The ionosphere is characterised by its variability and inhomogeneity. One of the characteristic inhomogeneities is the so-called auroral oval, which determines the range of auroral radiance. Detection of the auroral oval is an important task for forecasting auroral storms, as they affect long-range communication systems, navigation, satellite-to-ground communications, making communications complicated or impossible. Therefore, an auroral oval detection and prediction needs to be performed in order to be informed about the area of their possible influence at certain time intervals. On the basis of the available image dataset from SIMuRG, which is based on GNSS data, it is proposed to use the LSTM model and CNN architecture. The paper reviews existing implementations and proposes a method for predicting auroral oval movements in the images, using the Convolutional LSTM architecture, which combines time series processing and computer vision. The work results in a machine learning model that can make the predictions based on even small sets of data.

Everyone always wants to have more money. There is a race going on around the world, and the richest is the winner. There are different ways to make money, doing a 9-5 job, starting a business, investing in a business. Some new ways are the promoting different brands online for some return. Making money is the goal of new digital era. One of the ways to make money is to invest in companies by buying their stocks/ shares and then getting profit from the companies based on that year’s Fiscal performance. Trading can be done through the stock exchanges across the world. A person or even company (legally considered a person) can buy stocks of an entity based on their own predictions and thoughts whether the stocks will give some sort of profit or not.

Research on fire detection has grown steadily over the last few decades which is the key concern of the research community to prevent the lives of mankind and their property from damages. Several researchers developed video fire detection datasets and proposed different machine learning algorithms for its accurate detection. Therefore, it is very significant for the researchers to understand the relevant datasets in this field that can provide help in terms of results comparison and speed up the research based on the existing datasets instead of creating a new dataset. In this paper, we provide a comprehensive overview of existing fire detection datasets. Firstly, we reviewed seven different fire detection datasets in detail, which can provide helps to new researchers in this field. Secondly, we provided a detailed description of these datasets, and analyzed the shortcomings and suggestions for further fruitful research. This paper is helpful for new researchers to identify possible research gaps and limitations about fire detection datasets.

With advanced internet of things (IoT) and cloud/edge computing, industrial control systems (ICSs) are evolving. However, there are critical concerns and challenges about the cybersecurity of the IoT-enabled ICSs against cyber-attacks. To reduce the risk of cyber-attacks, an intrusion detection system (IDS) is required. In general, IDS utilizes signature-based or behavior-based methods to detect potential harmful anomalies. In this paper, we propose a hybrid intrusion detection approach deploying a statistical filtering method and a composite autoencoder to effectively detect anomalous behaviors caused by cyber-attacks. The proposed method is validated by experimental data acquired from a real water treatment system as a case study of cyberattack on ICSs.

Recently, cloud computing technology has been rapidly growing faster, offering cloud-based human activity recognition applications with lower latency. In this paper, we design and implement a new distributed driver activity recognition system (DHAR). The proposed distributed system absorbs a more significant number of input sensor data from humans with a lightweight model that provides high accuracy for driver activity recognition. In addition, our model has employed the entire convolution network – Long Short-term Memory (FCN-LSTM) to predict human activities of a total of 6 classes such as walking, walking upstairs, walking-downstairs, sitting, standing, and laying. We evaluate the proposed system using a well-known UCI-HAR opensource dataset containing a collection of smart-phones data for 30-subjects while performing various activities using a smartphone. We used various Amazon cloud computing services for the deployment of the proposed architecture. The experimental results show that the proposed architecture improves end-to-end latency by 2.7 times compared to the traditional architecture.

The fields to which IoT applies are widely found in domestic, industrial, national, and other facilities, where it serves various functions depending on the environment. However, most IoT systems communicate with servers and control IoT devices to function according to the server's requests and in this process, there is the threat of an IoT attack. Studies on security such as authentication systems as communication security techniques are undertaken to cope with it. However, systems with limited resources, such as IoT, require overhead to perform encryption operations for authentication. To solve this problem, the mechanism proposed in this paper performs authentication through subchannels and injects specific patterns into time stamps and time intervals to authenticate them. The mechanism does not use additional authentication solutions in limited hardware resource environments and performs authentication with physical timers and interval operations inside the system.

According to security breach level index millions of records are stolen world widely. Keeping in view the application of blockchain we have proposed novel secure mechanism based on neural network to train the model and to record the behavior of the users accessing the personal health records through blockchain. Recent solutions to storing and distributing patient data have several confines that limit users’ access to their patient health records (PHR), decrease the accessibility of critical information to care providers, and eventually extant a boundary in the transfer of traditional healthcare a digital healthcare approach. In order to remediate such shortcomings, a blockchain-based solutions the best tool to provide storage and trust. Numerous cloud-based approaches have been suggested in digital healthcare data allocation, but it would be untrustworthy to rely on a third party. In recent times, blockchain has been implemented in digital healthcare record sharing, which skip trusting on a third-party.

Now a day’s social media become a trend for the promoting marketing. It provides a platform where people can share or promote their ideas, business, and products as well services. The promoting campaigns are according to the psychology of user in modern world. This discovery is useful for the people who use these sites for the business promotion. It provides a platform for the novice as well as the expert users to develop their business. They can achieve their aim by making different pages for the business promotions. The purpose of this research is to explore which social media site is best for the development of the business. The comparison between Facebook and Twitter is done along with their usability investigation.

Care and nursing training (CNT) is to develop the ability to effectively respond to the needs by investigating patients' requests and improving trainee' care skills in a caring environment. Although conventional CNT program has been conducted based on videos, books, and role-playing, the best way is to practice on an actual human. However, it is challenging to recruit patients for training continually, and the patients may have experienced fatigue or boredom with iterative testing. As an alternative approach, a patient robot that reproduces various human diseases and provides feedback to trainees has been introduced. This study presents a patient robot that can express feelings of pain states like an actual human does in joint care education. The primary two objectives for the proposed patient robot-based care training system are (a) to infer the pain felt by the patient robot and intuitively provide the trainee with the patient's pain state, and (b) to provide the facial expressionbased visual feedback method of the patient robot for care training.

There are many openly available medical resources, which include structural clinical and genomic data for cancer patients. The data sources are provided with primitive toolsets which provide minimal descriptive statistics to help in initial exploratory analysis. However, this minimal support is not sufficient in many ways. Firstly, with an individual data source, the tools are not supporting statistical models and visualization to infer preferable clinical association with gnomically mutated patient data. Secondly, these data tools are not coherent to integrate the external toolsets for sharing the data and analytical outcomes. To target this limitation, the paper provides a technical detail of the development of Webbased tools – known as BioKG portal (Biological Knowledge Graph). At this moment, the BioKG is coming with features; a) provide unified data-model (influenced from knowledge graph), b) integration of different data sources with propriety data support – such as CSV files, JSON based API, other custom delimited files, c) integrated gene set enrichment analysis support on unified clinical and genomic data, and d) visualization support for the analytical outcomes. The BioKG portal is demonstrated in thyroid cancer by integrating data from openly available portals – TCGA and cBioPortal. This web-based toolset's key advantages are facilitating stakeholders with diverse capabilities – such as clinicians, bioinformaticians, and computer scientists.

Heart illness is a serious disease faced by a significant populace worldwide. Considering death rates and large numbers of patients with cardiac illness, it is shown how important it is for a person to be subject for early diagnosis of the disease, as an average citizen cannot afford frequent expensive tests such as the ECG. So, an effective system needs to be put in place that is both practical and reliable for the prediction of the chances of heart disease. The extraction of medical data has become highly recommended so that the rate of death which is very high due to heart diseases can be predicted and treated. The development of a machine-based system for cardiac diagnosis provides a more accurate diagnosis than traditional methods. We therefore propose an app to prevent the vulnerability of heart disease given the fundamental symptoms, such as age, gender and rate of the pulse and so on. The neural networks of the machine learning algorithm have proven to be the most reliable and accurate algorithm used for this system.

The use of surface electromyography has increase recently for hand gesture recognition because of the feasible usage of low cost, wearable, non-invasive devices. Hand gesture enhances human-machine interaction to great extent. This paper proposed a robust approach for hand gesture classification using various machine learning classifiers. Six different features such as; minimum, maximum, peak to peak, root mean square, zero crossing and waveform length are extracted from raw data and fed to machine learning classifiers. Data is comprised of 36 individuals and seven gestures are classified with an accuracy of 90% and F1 score of 87% using Support Vector Machine classifier. Our reproducible implementation is available at github.com/talhaanwarch/emg-gesture-classification

MRI imaging is expected to play a key role in diagnosing and arranging brain tissue treatment. It helps specialists in determining the stage of a brain tumor previously. Because of the complicated brain anatomy, MRI is used to diagnose brain tumors. Imaging, which is complex and a challenge. MRI images allow for greater differentiation between the soft tissues of the human body. MRI images skipped CT scans, ultrasound, and X-ray results. MATLAB image processing (IP) tools are used to perform pre-screening, background processing, and methods such as Filtering, contrast brightness, Edge location and background management processes for the receiving of brain tumor images, such as Histogram, Threshold, Segmentation, Morphological activity.

In late 2019 a new pandemic by the name of COVID was detected in the world. This pandemic, which started in China, spread to almost all parts of the world. It is reported that 226,411,199 people are infected by it, and the graph is going up every day. The researchers identify the leading cause of this spread is the human-to-human interaction. Governments have enforced stick restrictions on the movement of people that helps to reduce the spread of the virus. Google COVID-19 community mobility reports is assisting the researcher in the community to provide the human mobility statistics during this pandemic. This research focuses on gathering some answers which can help humanity in fighting against the virus. This research analyzes all the mobility during this pandemic time, which will help us analyze the spread of COVID-19, which will help to take control measures to stop the spread of COVID-19 in the country.

Compaction is an essential ingredient in a LSM (Log-Structured Merge)-tree based key-value store. In this paper, we analyze two representative compaction policies, called leveled and universal, using RocksDB. Our analysis uncovers that the universal policy has a capability to reduce write amplification by applying compaction in a lazy manner. However, the lazy manner deteriorates space amplification, which leads to adverse effects such as a relatively longer period of low performance for compaction and degraded throughput for range query. We also observe that, for sequential access pattern, the leveled policy can provide better write amplification than the universal policy by employing a technique called trivial move. In addition, we find out that the background compaction and index cache give a substantial impact on the performance of point query. Our analysis reveals tradeoffs between two policies based on various aspects including access pattern, query type, and configurations, which can be used effectively for designing new and hybrid compaction policies.

Recently, with the development of IoT technology, data has increased, and problems with the centralized cloud computing method are appearing. As an alternative to this, Edge Computing, a distributed cloud method that processes data close to the edge of the network where data is generated, is utilized. On the other hand, as the number of containers on one host increases, container management becomes difficult, and accordingly, container orchestration technology capable of configuring and managing a large number of containers is required. In this paper, we measure, compare, and analyze the time to transmit sensor data to the DB server of each Kubernetes cluster by building a multicluster infrastructure using Kubernetes, which is the most used container orchestration tool.

Renewable energies use clean sources for energy generation and have the potential to balance the supply and demand of power. One of the best ways to save energy for high-demand time is to preserve it in a battery energy storage system (BESS). Various methods are presented in the last two decades for battery state of charge (SOC) estimation, however, most of them are focused only on a single battery pack and use data without accurate preprocessing and feature selection strategy. Therefore, in this paper, we conduct a comparative analysis of machine learning (ML) models with a specific preprocessing strategy and suggest a high performer model for battery rack SOC estimation. First, we preprocess the data by cleaning, normalizing, selecting important attributes, and then split it into training and testing sets. Next, four ML models are trained using the training data for SOC estimation, and finally, for better evaluation, each model is evaluated on the testing data using various error metrics. After comprehensive experiments, we suggest multilayer perceptron (MLP) due to high performance for batteries rack SOC estimation.

Recently, with the development of IT technology, interest in technology using bio-signals is increasing, and biometric authentication technology using a part of the human body or behavioral characteristics are widely used in communication, automobile, security, medical care, corporate marketing, and public fields. However, as the importance of security increases, the awareness of security is being reevaluated, and there are security problems against the vulnerability of attacks on information transmitted in all transactions and forgery and falsification of existing biometric authentication. Accordingly, recently, research on a method for authenticating a user using two or more different authentication factors has been actively conducted. In this study, we propose to develop a multi-bio authentication system that guarantees increased convenience and high reliability through the development of a security-strong system with bio-signal-based next-generation bio-authentication.

In this paper, we selected 15 countries for comparative analysis of total and individual types of alcohol consumption and among which, 7 countries of Life Expectancy (LE) from both developed countries and developing countries. We categorized alcohol drinks into three types, namely, beer, spirit and wine. Interestingly, although some countries have very similar amount of alcohol consumption per capita, the result which led by alcohol consumption of different types can make a big difference. In other words, the impact of different types of alcohol is unequable.