Advances in Data-Driven Computing and Intelligent Systems

Ultra-high-performance fiber-reinforced concrete (UHPFRC) is a new class of concrete material that produces superior compressive strength (more than 150 MPa). In addition, UHPFRC shows significant improvements in post-cracking strength, strain capacity, and specific work to fracture based on the strain hardening behavior under tension. Thus, UHPFRC is expected as a potential material for civil infrastructure although several of its properties have not been published yet. This study aims to evaluate influences of specimen and fiber sizes on the tensile performance of UHPFRCs. Two sizes of specimen and two sizes of smooth and twisted fibers were used in the production of UHPFRCs. From the experimental test results, the tensile behavior of UHPFRC could change from strain hardening behavior to softening behavior as the specimen size increased, whereas the increase of fiber size did not alter the strain hardening behavior of UHPFRC. The tensile performance of UHPFRC was highly dependent on the fiber size and specimen size. The post-cracking strength increased, but the strain capacity significantly reduced with increasing the specimen size. On the contrary, the post-cracking strength decreased, but the strain capacity marginally enhanced with increasing the fiber sizes.

The study deals with modeling and automation of obtaining, collecting and analyzing health data in smart medical wards. The authors made a review of various existing solutions in the domain of interest. Such solutions increase safety and comfort of the patients, enhance quality of their treatment in hospitals, reduce the costs and the influence of human errors. However, the existing solutions are aimed at solving particular tasks. The use of different tools and heterogeneous modules causes problems of compatibility, security risks and increased complexity and redundancy of systems, so they become more expensive to support and maintain. To overcome these obstacles, the smart ward can be built within a general framework that supports typical modules, interfaces and operations. At the first stage of solving the problem the authors proposed a conceptual model of a smart ward, and make a description of its components and operation scenarios. Then, the authors considered a sample scenario for disease stratification based on obtaining values of health indicators and performing calculations according to a defined methodology. An algorithm for data collection and analysis for diagnosis making and disease stratification was developed, and a simulation example of smart ward functioning was carried out. Numerical experiments demonstrated that data collection and processing using smart components of the model could significantly reduce the assessment time.

Parts-of-speech (POS) tagging is considered one of the most challenging fields in natural language processing (NLP). The objective of this research is to develop a POS tagger for the Assamese language. Due to the scarcity of digital linguistic resources, Assamese lacks high-performing POS taggers. To fill this gap, long short-term memory (LSTM) and bidirectional long short-term memory (Bi-LSTM) are explored in the proposed research to develop a POS tagger using an Assamese POS corpus. It is important to note that this experiment faced difficulties in understanding and managing natural language for computational linguistics, which was also anticipated. The Assamese corpus considered in this research comprises around 50,000 words. At the initial stage, while examining the first few sets of data, which are about 20,000 words, it was noticed that the taggers yielded satisfactory results. Based on the result derived, the Assamese corpus size has been enhanced to 50,000 words and better performance is noted in terms of accuracy rate, precision, recall, and F1-score. As a result, an accuracy of 91.20% is achieved for LSTM and 91.72% for Bi-LSTM. Concerned with substantial research from the NLP perspective for the Assamese language and for comparative purposes, a comparison between existing POS taggers in the Assamese language and the proposed work is also presented.

Detecting explicit lyrics in Hindi music is a crucial responsibility to prevent the public from hearing offensive and improper material. In recent years, different machine learning methods have been used to find obscene lyrics in Hindi music. K-Nearest Neighbours (KNN), Convolutional Neural Networks (CNN), Naive Bayes, and Long Short-Term Memory (LSTM) are four techniques for recognising explicit lyrics in Hindi music that are compared in the present research. Hindi song lyrics that have been classified as explicit or not make up the dataset used in this study. Python is used to implement each method, and multiple performance metrics including accuracy and F1-score are used for evaluations. The findings demonstrate that, with an accuracy of 81.6% and an F1-score of 0.96, the LSTM model surpasses every other method. To see how the CNN and LSTM models learn, the training and validation graphs for each model are shown. The graphs show that the LSTM model performs better at detecting explicit lyrics in Hindi music than other machine learning models, with greater accuracy and lower loss. Overall, the study underlines the benefits of employing LSTM over alternative approaches and shows how well machine learning methodologies work for identifying explicit lyrics in Hindi music.

Multimedia gaming for learning and development denotes an enduring immersive design and effective learning challenge for academics, students, researchers, industry stakeholders and software developers globally. Inconsistent evidence for gameplay learning effects compared with traditional learning delivery tends to bear the “learning gaming puzzle” continuing. Sparse findings grapple with open innovation-related aspects and a less clear impact is noted for global citizenship influencing sustainability. Thereby, this research expands on assessing open innovation and civic learning properties in conventional and learning gameplay context in a food industry. We elaborate on 65 Greek employees’ open innovation and civic capacity attributes assessed post-workshop and after-Resilience gaming instruction in 2022. The employees did perceive the Resilience game as more favourable instructional tool for open innovation and civic agency in relation with traditional learning mode. They further related the Resilience game to direct open innovation-associated elements in their organization. The acquired findings are discussed in light of theoretical and hands-on ramifications and routes for additional exploration regarding learning game open innovation and sustainability-linked L&D and experience.

Epileptic seizures is a well-known and the most chronic neurological disorder and it is observed highly in infants and elderly people. Symptoms of epileptic seizures are very dangerous; they lead to injuries due to falling because of jerking movements of the arms and legs that can’t be controlled. Electroencephalography (EEG) signals are prominent tools to analyze brain activities and detect seizures. Learning machines are promising in detecting seizures by analyzing EEG signals. This study proposes a customized deep learning network for seizure detection, i.e., DLN-SD utilizing EEG signals. The publicly available CHB-MIT Scalp EEG dataset is considered. It is empirically compared with pre-trained models, namely AlexNet and GoogleNet and observed that the proposed model (DLN-SD) brings 96.0% precision, 98.0% recall, and 98.0% accuracy and performs best among all three models. It is concluded that the DLN-SD shows an improvement of 15% over AlexNet and 9% over GoogleNet to detect seizures from the EEG dataset.

Modeling anomalous behavior pattern has become as a significant research domain in the recent years due to the security demands in public places. In literature some of the existing approaches such as statistical-based, density-based are applied for pattern detection whereas traditional approaches may not suitable for all scenarios, since they are limited with their properties. In this study, we propose deep learning-based behavior recognition to characterize the abnormality through train and test the frame. Proposed system used the AlexNet Convolution Neural Networks (CNN) architecture, which has largely trained dataset used for feature extraction. In addition, optical flow applied to estimate the human motion. Moreover, motion influence map utilized to reflects the characteristics of the human motion. The CNN-based proposed system can steadily outperforms abnormal behavior detection, especially for the lightweight models when given a small among of training samples.

It is challenging and requires automatic recognition of handwritten letters and numbers, especially in the era of digitalization. However, several literatures exist in major and nationalized languages. In the last few years, in an era of the paperless office, it has been necessary to cover all the most utilized languages in the region. Optical character recognition (OCR) with deep learning has achieved remarkable performance in various character recognition and interpretation tasks. In this study, the authors proposed a vision transformer (ViT) base approach (GujFormer) to recognize Gujarati handwritten characters. The proposed study used a multihead self-attention model to enhance feature learning. The proposed methodology used three different datasets having handwritten characters, numbers, and printed Gujarati characters. The vision transformer achieved 98.31% accuracy for handwritten digits and 97.92% accuracy for handwritten characters. The methodology used multilayer perceptron (MLP) as a classification layer. Simulation of the proposed methodology was also compared with other deep learning CNN models like VGG16, InceptionV3, DenseNet121, and others. Authors have also analyzed recognition variation over different self-attention heads. The proposed methodology has achieved remarkable performance over vision transformer for Gujarat handwritten character recognition.

Farmers can benefit from the knowledge of soil characteristics to implement efficient and sustainable agricultural practices which yield more crops with less use of resources. Using a machine learning technique, research is attempting to predict soil characteristics. Sand, pH value, Soil organic Carbon, Calcium, and Phosphorus are the primary characteristics of soil forecasts. These characteristics have a significant influence on crop yields. Six commonly used machine learning models are used for the prediction of these soil characteristics: Linear Regression, Random Forest Regression, Support Vector Machines, Gradient Boosting, XGBoost, and AdaBoost. The performance of these models is accessed by reference to the African Soil Property Prediction Dataset. The prediction performance of each model was evaluated using mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination (R2 score). The results suggested that among all the evaluated models, the XGBoost Regressor model was the most suitable prediction model for soil properties.

In this paper, we propose an automated waste classification system that uses the help of deep learning and image processing methods. The main objective of our proposed system is to streamline the classification of waste into organic and recyclable. To achieve this, we used image processing techniques to improve the image quality of recyclable and organic waste from a publicly available dataset and trained our model using convolutional neural networks (CNN). The model achieved a high validation accuracy of 90.65%. We compared this model with four pre-trained models based on their accuracy and loss. Our proposed model for automating waste classification holds the potential to improve waste management practices and can also reduce the amount of waste that ends up in landfills.

Food safety is an important issue nowadays. Now, people become more aware of food safety and health concerns as their lifestyles have changed they follow more precautions for food consumption. The conventional technique of food safety is not capable of ensuring food safety very precisely and has many limitations so the food business operators are facing economic losses. To resolve these limitations, the researchers and scientists developed AI-based integrated machine learning technology. Machine learning is a computational science that enables learning, reasoning, and decision-making by interpreting and analyzing statistical patterns and subject data. ML tools like sensors, electric nose, and tongue, smart indicators, line-scan hyperspectral imaging systems, RT-PCR, ELISA, etc. are capable of overcoming these limitations of food safety practices by using the ML algorithms. The ML technique has great potential to overcome food safety issues and future applications in the food sector.

The process of digital transformation exposes organizations to cybersecurity threats and numerous vulnerabilities. As these threats and attacks become increasingly sophisticated, the need for effective security measures to protect valuable assets and maintain confidence in the digital environment has become even more critical. Adopting an Information Security Management System (ISMS) assists organizations in achieving effective governance in relation to information security and business continuity. Among the numerous security standards available (e.g., NIST CSF, COBIT, and PCI DSS), ISO 27001 is one of the most highly adopted security frameworks. Security standards differ from each others in their requirements, policies, and best practices, which makes it the responsibility of decision-makers to choose the appropriate one based on the specific needs and objectives of their organizations. This paper aims to provide an overview of the structure and components of ISO 27001 and how it evolves over time. Furthermore, a comparison between ISO 27001 and two other popular security standards, NIST CSF and COBIT, is provided.

Predicting potato and tomato leaf disease is vital to global food security and economic stability. Potatoes and tomatoes are among the most important staple crops, providing essential nutrition to millions worldwide. However, many tomato and potato leaf diseases can seriously reduce food productivity and yields. We are proposing a hybrid deep learning model that combines optimized CNN (OCNN) and optimized LSTM (OLSTM). The weight values of LSTM and CNN models are optimized using the modified raindrop optimization (MRDO) algorithm and the modified shark smell optimization (MSSO) algorithm, respectively. The OCNN model is used to extract potato leaf image features and then fed into the OLSTM model, which handles data sequences and captures temporal dependencies from the extracted features. Precision, recall, F1-score, and accuracy metrics are used to analyze the output of the proposed OCNN-OLSTM model. The experimental performance is compared without optimizing the CNN-LSTM model, individual CNN and LSTM models, and existing MobileNet and ResNet50 models. The presented model results are compared with existing available work. We have received an accuracy of 99.25% potato and 99.31% for tomato.

Nowadays, manipulated images can be found everywhere due to the wide availability of sophisticated image editing tools. As the digital technology advances for multimedia, the field of image forensic has also made significant progress in recent years with the development of effective and more advanced image tampering detection methods. Furthermore, there has been continuous development in the image tampering methods along with the techniques used for hiding tampering clues aiming to evade the existing forensic approaches. Image splicing is one of the most common and simple image tampering approach, where an object or region is copied from one image and stitched into the other image. This paper presents an anti-forensic analysis on three advanced filters: the weighted average filter, bilateral blur filter, and Kuwahara filter. The objective of this analysis is to examine the impact of filtering operations on the detection accuracy of image splicing. Two deep learning-based image splicing detection models are proposed which are initialized with pre-trained weights of ResNet and InceptionNet architecture. To assess the effect of filtering, the structural similarity index measure (SSIM) is employed to quantify the similarity between images before and after the application of the filters. Based on the experimental results, the weighted average filter and Kuwahara filter emerge as the most effective anti-forensic attack on spliced images. These filters demonstrate the ability to preserve the SSIM of the image while inducing a substantial decrease in the accuracy of splicing detection.

In this research work the resonant frequencies and the corresponding mode shapes of circular Chladni plate is used to analyze the vibration response of the plate. A camera captures Chladni patterns for training, classification, and prediction of respective frequencies using deep learning algorithm. The Chladni patterns formed, depend on the material properties, geometry, thickness, and the vibration frequency. The experimental setup consists of a circular Chladni plate of acrylic with 300 mm diameter and 2.48 mm thickness. The simulation of mode shapes software ANSYS (student’s version) is used. Comparison of the Chladni plate experimental results and ANSYS simulation shows error of < 14%. Deep learning techniques shows the validation accuracy with pretrained network VGG16 as 99.16% and 99.49% with GoogleNet and prediction accuracies 98.63% with GoogleNet and 97.98% with VGG16. The resonant frequencies along with its mode shapes can be predicted using the Chladni plate approach combined with deep learning.

Stress is a concern in today’s paced society impacting individuals in various aspects of their lives including educational environments. It is crucial to identify and examine students’ stress levels as it offers insights into their well-being academic performance and excellence of life. The goal of this study is to develop a mental stress detection system that utilizes a combination of multiple sensor’s data fusion and deep machine learning (ML) models with the help of three physiological signals named electrocardiogram (ECG), galvanic skin response (GSR) and skin temperature (ST). Data is gathered from 200 students with the help of some stressors using a novel Internet of Medical Things (IoMT) device developed using low-cost sensors. Then the data went through pre-processed methods before being analyzed using some techniques of deep ML models like multi-layer perceptron neural networks (MLPNN), recurrent neural networks (RNN), and long short-term memory (LSTM) networks to identify stress. The results show that the RNN and LSTM models regularly outperform the MLPNN model in recognizing mental stress. These models adequately depicted the dynamic character of stress reactions while efficiently capturing temporal dependencies. The accuracy of the RNN and LSTM models was 92.34% and 93.20%, respectively. These models’ improved performance demonstrates their potential in stress detection applications. The combination of multi-sensor data fusion with deep ML algorithms allows for the precise detection of stress-related physiological changes. This has major consequences for student well-being and academic success since it enables personalized stress management tactics and interventions. Student’s mental health and academic experience can be improved by proactive efforts.

Skin diseases encompass a wide range of conditions, from minor issues like acne and hives to severe ailments like eczema, psoriasis, and melanoma. Deep neural networks have been effective in classifying these disorders, but they often exhibit an ethnic gap, leading to lower detection accuracy for darker skin complexions. This paper demonstrates a novel approach making use of the vision transformer network by combining the original and binary segmentation masks using attention to minimize bias and classify skin diseases. The proposed methodology involves preprocessing images, generating segmentation masks, fine-tuning the vision transformer model, and evaluating performance using metrics like accuracy, precision, recall, and F1-score. The model achieved an accuracy rate of 91.3% when trained and tested on three different skin disease datasets. The experiment’s results indicate the proposed method’s effectiveness in classifying diseases of skin lesions belonging to the different ordinals of the Fitzpatrick scale and highlight the potential of further study in this field.

The advent of multimedia-based communication requires high-speed, memory-efficient data storage of high quality images. One method for reducing the dimensions of an image while preserving the finer texture characteristics is image super resolution. This paper proposes a new model to deal with the reconstruction time and artifacts that GAN-based deep learning models face. To increase the speed, FASRGAN aims at reducing the complexity of generators. The network architecture builds on the ideas from ESRGAN and RAMS to derive a more efficient FASRGAN. This paper introduces the Residual Feature Attention Block (RRFAB) as a fundamental feature extraction block. Moreover, a relativistic discriminator is employed, inspired by relativistic GAN, that predicts the realness value of the generated image rather than a strict class value. Ultimately, training is performed on the model to improve on the content loss, which helps to converge the weights in the later training stages. The proposed model FASRGAN, achieves PSNR scores 31.5 and 31.04, SSIM scores 0.975 and 0.93 on Set5 and Set14 datasets, respectively, that are greater than the state of the art techniques and SSIM scores that are comparable to them.

This paper presents application of machine learning and geospatial analysis to examine sentiments from Twitter data during the 2023 season of the Indian Premier League (IPL). Utilizing machine learning models (linear SVM and logistic regression), we effectively categorized sentiments into negative, neutral, and positive classes, attaining impressive overall accuracy of 96.8% and 97.1%, respectively. The geospatial analysis led to map sentiment across various geographical locations in India, reflecting diverse public sentiment throughout the IPL season. We noticed intriguing temporal and spatial variations in sentiment distribution across March, April, and May in 2023, visually represented via a heatmap. This comprehensive examination of sentiment distribution, linked to a significant event like the IPL, opens new horizons for understanding public perception and emotional response to large-scale sporting events.

Ship classification is essential in coastal areas to ensure safety, protect the environment and improve maritime security. It also allows the optimization of resource allocation and can boost economic growth. Therefore, vessel identification is crucial to employ appropriate security measures. However, communication interruptions can happen during poor weather conditions, which could hinder the overall safety of vessels in the area. Security is, therefore, a main pivotal aspect that drives forward the vessel identification systems. This paper tackles this problem by proposing an XGBoost machine learning model that is optimized by an enhanced variant of the sine cosine metaheuristic algorithm that has the role of identifying and classifying naval vessels. The proposed method has been compared to the results attained by other cutting-edge metaheuristics algorithms, and experimental outcomes show that it obtained supreme results for this particular task.

The application of technology for sensing, analysis, control, and communication within ground transportation is referred to as an intelligent transportation system. This system aims to enhance safety, mobility, and efficiency. Intelligent Transportation Systems (ITSs) are in the process of development and implementation, leading to improved accuracy in predicting traffic flow. The efficacy of traveler information systems, public transportation, and advanced traffic control is said to depend on these systems. In order to effectively manage and lessen traffic congestion, practical execution is essential, as evidenced by the expanding use of data in transportation management. By employing machine learning (ML), it is possible to construct predictive models that incorporate diverse data from numerous sources. Predicting traffic movement, reducing congestion, and identifying optimal routes that consume the least time or energy all require traffic prediction, which involves forecasting traffic volume and density. Traffic estimation and prediction systems have the potential to reduce travel times and enhance traffic conditions by enabling more efficient utilization of available capacity.

Due to rapid advancement in electronic commerce technology, the use of credit cards has dramatically increased. As credit card becomes the most popular mode of payment for both online as well as regular purchase, cases of fraud associated with them are also rising. For many years, countless supervised machine learning approaches for anomaly detection have been proposed that have achieved state-of-the-art performance. For our deep reinforcement learning agent, we describe a revolutionary deep Q-network architecture and a customized OpenAI Gym environment in this paper. This design makes use of experience reply and is based on value function approximation. In order to carry out classification procedures determined by batches of input data; the deep Q-agent makes use of an epsilon-greedy policy. Subsequently, the agent is evaluated and rewarded by the OpenAI environment based on its performance in accordance with the assessment. The recollection of the agent contains every single detail of this adventure. The deep Q-agent takes several memory samples from its experience buffer at the end of the batch completing process. It then uses the Q-network to update the Q-value and the loss is calculated. Further, the weights are improved using backpropagation. The proposed procedure has achieved a level of performance that is state-of-the-art, and the results reveal that it was effective in correctly identifying fraudulent and non-fraudulent transactions.

Wireless sensor networks (WSNs) are created by deploying various sensor units in harsh environments for target monitoring, as well as by deploying various tiny sensors collectively. The sensor units monitor the data in their vicinity and send it to the sink for processing. However, due to various parameters such as coverage area, in-built sensing and processing units, and so on, the sensor nodes may consume more power when forwarding information to the sink. We concentrate on the existing issues in the coverage control of various sensor models for quantitatively and qualitatively solving the sensor network coverage area. Various optimization techniques such as spotted hyena optimization (SHO), particle swarm optimization (PSO), and moth flame optimization (MFO) are studied and discussed. The trade-off between coverage rate, total number of sensor nodes, and energy consumption was observed, and the techniques’ significance was compared.

Sparse optimization problems at a large scale present considerable difficulties in diverse fields, such as machine learning, data mining, and signal processing. The aim is to identify the most efficient solutions within expansive search spaces, while constraining the number of non-zero variables. The present study aims to tackle this obstacle by introducing an innovative methodology that integrates whale swarm optimization (WSO) with neuro-evolutionary computing (NEC). The WSO algorithm draws inspiration from the social behavior of humpback whales and presents a proficient approach for investigating and utilizing the search space. The integration of evolutionary algorithms with neural networks by NEC has demonstrated potential in effectively addressing intricate problems. Through the simultaneous evolution of neural networks and optimization processes, the algorithm is capable of adapting and enhancing its search strategy progressively. The study aims to improve the efficiency of WSO in addressing optimization problems characterized by sparsity at a significant scale by incorporating techniques from neuro-evolutionary computing. The method under consideration entails the integration of neural networks into the WSO algorithm with the aim of directing and augmenting the exploration procedure. The study evaluates the efficacy of the hybrid algorithm in relation to established methodologies based on criteria such as the quality of solutions, rate of convergence, and computational efficiency. The results obtained from the experiment provide evidence to support the efficacy of the proposed approach. The hybrid algorithm effectively utilizes the sparsity structure inherent in the problems, resulting in expedited convergence and enhanced precision of the solutions.

The traffic issue in Bangladesh is one of the strongest and most demanding issues in city surveillance today. Finding and separating automobiles on the side of the road and in the parking lot has become more crucial as Bangladesh’s traffic congestion problem worsens at an alarming rate. There is now a lot of study being done on the topic of object detection and classification. The detection of vehicles and the recognition of license plates have been the subject of numerous studies. But there are still certain restrictions. Locating the double-row number plate accurately is one of them. In this study, we proposed a technique for locating both single-row and double-row license plates as well as detecting the vehicle type. The suggested model, YOLOv4 and OCR (optical character recognition) Tesseract can be utilized to create a real-time system and has good accuracy and inference time for a variety of illumination and style of Bangladeshi number plates. The model showed a mAP value of almost 97%, and the other evaluation metrics performance is also acceptable. Our proposed model outperformed the prior system.

Tip burn is a common calcium deficiency symptom found in strawberry leaves, which can deform, fold over, and discolor the tips or edges of budding leaves. To help farmers detect diseases early and manage their crops more effectively, this work aims to build a solution that employs image processing and machine learning techniques to detect tip burn in strawberry leaves on a farm. The proposed work utilizes a dataset of 1431 images of ‘diseased’ and ‘healthy’ strawberry leaves. A Random Forest classifier, which is an ensemble learning model built on decision trees, CNN, or convolutional neural networks, and VGG-16, a transfer learning-based approach were all fitted in the PySpark environment to assess the feasibility and scalability of using Spark for distributed computing. The results indicate the deep learning approaches outperformed the traditional machine learning approach in accuracy and execution speed, making them a promising tool for identifying tip burn diseases in strawberry leaves. Overall, this work provides farmers with a fast, feasible, and reliable method for detecting tip burn, aiding in crop management, and reducing crop loss.

Plant diseases reduce yields, directly affecting domestic and global food production systems. Using image classification and early prediction of plant diseases can help us to manage yield production properly. This study evaluated the deep learning models VGG19 and ResNet50 for potato leaf disease classification and prediction. The performance of deep learning models VGG19 and Resnet50 is recorded based on performance metrics such as confusion matrix, precision, recall, accuracy, f1-score, and ROC/AUC. The models VGG19 and Resnet50 achieved the highest accuracy at 93% and 92.58%, respectively.

Real-time small object detection in aerial images is a challenging task in computer vision because of the complex environmental conditions and fewer pixels occupied by the objects. Along with this getting real-time inference makes this much more difficult. To tackle these challenges we propose an improvement in the classical YOLOv3 network which is capable of removing noise from the images and performing small object detection to while maintaining a high frame frequency. We are using a drone to capture real-time footage from the AIT campus, which is further used for classifying objects like vehicles, pedestrians, etc. Our proposed model “AeroDetect” was able to beat various state-of-the-art models like YOLOv3 and YOLOv5 in various parameters by having higher mAP, and a lesser number of layers and parameters while maintaining a high frame rate which is crucial for our use case. We achieved mAP of 0.402 on the Visdrone dataset as compared to 0.4 and 0.38 by YOLOv3 and YOLOv5, respectively. On the VEDAI dataset, we got mAP of 0.632, whereas YOLOv3 and YOLOv5 achieved 0.578 and 0.604, respectively.

Action recognition is a fundamental task in computer vision with applications spanning surveillance, human–computer interaction, and video analysis. The work comprehensively explores action recognition techniques, integrating spatial and temporal information to enhance classification accuracy. The process is implemented using the KTH dataset, an action recognition benchmark dataset that is widely used. The proposed approach employs a VGG16 architecture to extract spatial features from input video frames, capturing appearance-related details. Simultaneously, calculating the optical flow between consecutive frames in the dynamic images is utilized to extract temporal features from optical flow images, representing the video’s motion dynamics. These spatial and temporal features are fused to create a comprehensive action representation. For classification, a Bidirectional Gated Recurrent Unit (BiGRU) is employed, where the fused features are fed into the GRU to model temporal dependencies and capture long-term contextual information. The evaluation results validate the effectiveness of the proposed methodology. Fusing spatial and temporal features and incorporating BiGRU improves classification accuracy compared to individually using spatial or temporal features. These findings highlight the importance of integrating appearance and motion dynamics in action recognition.

Gohr presented benchmark work in CRYPTO 2019, and for the first time, deep learning was successfully applied to mount a cryptanalytic attack that was claimed to be better than its classical counterpart. He implemented deep learning-based differential cryptanalysis against the block cipher SPECK32/64. Lu et al. in 2022 built related-key neural distinguishers (RKNDs) against two lightweight ciphers, SIMON and SIMECK. They achieved significant accuracy for different number of rounds in both ciphers. In this paper, we construct a related-key neural distinguisher against ultra-lightweight block cipher PRESENT. The developed distinguisher provides some significant observations for the reduced number of rounds for PRESENT cipher.

This study delves into strategies for enhancing energy efficiency in 5G and 6G networks, focusing on network optimization, radio access techniques, and management. It examines research articles to pinpoint important strategies. Among the notable optimizations are the comparison of the energy efficiency of deploying small cells in various microcell topologies, resource allocation strategies for wireless energy transmission, and dynamic base station napping to preserve power during low traffic. Hybrid beamforming (HBF) and adaptive sectorization are presented as ways to reduce energy consumption and boost network capacity. In order to save energy and increase throughput, network topology management techniques including route diversity and inactive base station modes are investigated. Although these methods have the potential to significantly reduce energy consumption and improve network performance, further study and development are necessary to maximize the benefits of energy-efficient wireless communications in the future.

Path planning problem, which involves finding the optimal path from origin to destination, is a challenging problem with broad applications such as autonomous vehicles, mobile robot navigation, and game artificial intelligence. In this paper, we propose an approach that brings together the benefits of high-level planning using Dijkstra’s algorithm and Lyapunov-based control system (LbCS) to address this problem. In our approach, the path planning problem is formulated in two parts. Firstly, it involves an optimization task at a high level, with the objective function representing the overall path cost or distance via landmarks. Following this, a LbCS navigates between these landmarks, ensuring stability and convergence to each point in the sequence. Our approach offers a comprehensive solution to the path planning problem, striking a balance between path efficiency and computational feasibility. This research paves the way for advanced landmark navigation and path planning algorithms, extending potential applications to include both static and dynamic obstacle navigation during path planning. The effectiveness and novelty of our proposed method will be assessed through simulations and comparative analyses using LbCS with and without our proposed high-level planning. We anticipate our findings will significantly contribute to the current body of knowledge and future landmark navigation research. Our plan for future research is to extend this work to include obstacles and refine methodology to achieve better results.

Lung cancer is one of the leading causes of death nowadays. It has been found in older and younger people in recent years, and the figures are worrying. Detecting lung cancer in the early stage significantly increases the chances of survival. This research proposes a novel hybrid method for early lung cancer diagnosis. This approach uses the Tomek links method, the DBSMOTE algorithm for preprocessing, PCA for feature reduction, and the SVM algorithm for classification. Moreover, we analyzed five classification algorithms on the lung dataset: SVM, KNN, Naïve Bayes, random forest, and Rpart. We compared the classification results with the proposed approach in terms of precision, F1, recall, accuracy, and balance accuracy. The experimental results demonstrated that the proposed method attains the highest classification accuracy (96.72%) than other methods used for the experimental study.

Sound event detection (SED) aims to recognize and distinguish various types of events, encompassing those associated with humans, nature, the environment, household machinery, and more. In real-world scenarios, it is common to encounter multiple sound sources co-occurring, leading to overlapping sound events. For instance, such events could include bird songs, footsteps, or the engine sounds of motorbikes in realistic environments. SED finds practical applications across several fields, including Home Automation and Security, Multimedia Organization and Retrieval, and Human–Computer Interaction. In our research, we propose a model that leverages LSTM and CNN algorithms to identify different sounds automatically. In this model, to train and evaluate the sound samples, we utilized a Kaggle audio dataset which contains 18,873 files that were classified into 41 distinct classes. We extracted various audio data features from the samples, such as Mel-spectrograms, spectral bandwidth, spectral centroid, and statistical features like spectrogram mean and median. By combining these features, we successfully trained our LSTM-RNN and CNN models. In our research, LSTM achieved 73.1% accuracy on the chosen dataset.

Conjunctivitis is a common infectious ophthalmic ailment that is known to cause significant discomfort and even loss of vision in some cases. The distinction between ‘Normal,’ ‘Viral,’ and ‘Bacterial’ causes of ‘Pink eye’ is presently difficult and assessed through culture reports and is inferred using deep learning architecture. An artificial intelligence-assisted classification of conjunctivitis can be of importance and help in its early diagnosis. The present study aimed to propose an end-to-end, automated conjunctivitis classification using a pre-trained deep learning architecture called EfficientNet. A three-labeled classification of conjunctivitis for ‘Normal,’ ‘Viral,’ and ‘Bacterial’ was performed. The model was deployed in real time as a web page called ‘OpthaPredict’ and achieved accuracy and precision of up to 99%. The study contributes to the development of more precise and reliable methods for predicting the type of conjunctivitis in real-time. The implications of the present study extend to the roots of the Indian healthcare system, such as Auxiliary Nurse Midwives (ANMs), Accredited Social Health Activist (ASHA) workers, Primary Care Physicians (PCPs), and early career ophthalmologists who rely on accurate results for informed decision-making.

Due to miniaturization in electronic devices and cost reduction, Flying Adhoc NETwork (FANET) plays a predominant role in military and civil application domains. To reap the real benefits of FANET, congestion among the highly dynamic devices must be effectively controlled. The traditional Transmission Control Protocol (TCP) and its variants such as TCP Tahoe, TCO Reno, are not suitable for FANET due to the frequent changes in mobility, trajectory localization, traffic and node density. These techniques fix the same retransmission count irrespective of the current network condition. It unnecessarily increases the number of floating packets in the network which triggers further congestion. To mitigate this problem, Dynamic Retransmission Count Prediction (DRCP) algorithm is proposed in this paper to dynamically predict the retransmission count when the timeout timer expires. It utilizes the regression technique to predict the throughput and packet loss rate from the current network condition. With these predicted values and Machine Learning (ML) techniques, the proposed DRCP algorithm predicts the optimal number of retransmission count. The comprehensive simulation of the proposed DRCP with various regression and ML techniques is performed under varying network scenarios. The obtained results such as throughput, jitter, packet loss rate, and delay are compared with other TCP variants. From these analyses, it is shown that the proposed DRCP algorithm outperforms TCP Tahoe with static retransmission count and TCP.

Industrial production is the base of our modern society and provides in all fields sufficient products that make true the feeding of the world and the wealth of nations. This means, then, in the modern world we have lived in since Descartes and Galileo that we rely on science for production, hopefully constantly increasing efficiency and innovation. Up to now, no system has been found that guarantees success in always finding higher efficiency and innovations. In this paper, we look at the gap in systems theory, how production works from their standpoint, and try to extrapolate future strategies towards this discipline, and look at the question how this will affect efficiency and innovation, in principle, starting from this new knowledge and these insights. A significant result is the resulting strategy of how to design the path towards mass production and the rising role of Artificial Intelligence (AI) in production and society.

Numerous digital forensic technologies have been developed to investigate and prevent cybercrimes as they have grown in frequency. Social media forensics and mobile forensics are two of these techniques that have proven effective in the analysis of cybercrimes. Whereas mobile forensics gathers, analyzes, and interprets data from mobile devices, social media forensics gathers, analyzes, and interprets data from social media platforms. The integration of these two techniques has significantly improved cybercrime investigations by providing investigators with a wider range of digital evidence. The integration of mobile and social media forensics has enabled investigators to monitor a suspect’s activities on several platforms, such as social media websites and mobile devices. This link makes it possible for investigators to retrieve erased data from mobile devices, which is important evidence for cases involving criminal activity. Social media forensics has also helped identify suspects for investigations by looking at their online behavior and communication habits. Social media and mobile forensics working together have greatly improved the investigation of cybercrimes and given investigators access to a wider range of digital evidence. Because of this connectedness, detectives may now track suspects’ activities across several platforms and retrieve deleted data from mobile devices, both of which can be crucial pieces of evidence in cybercrime investigations.

Accurate prediction of chlorophyll-a is crucial for assessing eutrophication dynamics in reservoirs, lakes, estuaries, etc. This study focuses on developing and comparing two models (1. Artificial neural network, ANN; 2. Adaptive neuro-fuzzy inference system, ANFIS) for chlorophyll-a prediction in the Ashtamudi estuary, Kerala, India. Total Dissolved Solids, Turbidity, and Total phosphates were taken as input parameters for predicting chlorophyll-a for both ANN and ANFIS models considering their considerable influence on chlorophyll-a. The ANN model achieved a remarkable coefficient of determination (R2) of 1.0, indicating a perfect fit between the observed and predicted chlorophyll-a values. The performance of the ANFIS model was also good, with an R2 exceeding 0.95. In ANFIS, the membership functions, gbellmf, trimf, and gaussmf over-predicted the chlorophyll-a contents, while trapmf, gauss2mf, pimf, dsigmf, and psigmf under-predicted the chlorophyll-a contents. The ANN model performed better compared to the ANFIS model on chlorophyll-a prediction in the Ashtamudi estuary and can be used for future predictions of chlorophyll-a. The models generated can capture the complex relationships between turbidity, TDS, total phosphates, and chlorophyll-a concentrations. Moreover, these provide valuable insights into the temporal and spatial variations of chlorophyll-a concentrations, allowing for proactive management strategies to mitigate the impacts of eutrophication. The models generated find applications among water resource planners, environmental agencies, and policymakers.

Artificial Intelligence (AI) has become increasingly influential in our society due to its ability of mimicking human behavior and intelligence. AI-powered language models are one of the subfields of AI where the type of AI system uses deep learning algorithms to generate and understand natural language. The well-known AI-based language models include RoBERTa by Facebook AI, BERT, T5, ALBERT, and XLNet by Google, ELMo (Embeddings from Language Models) by AI2, GPT-3 by OpenAI, etc. AI-based language models possess the ability to generate novel and unprecedented outputs including realistic images, digital artwork, music, and written content. These outputs often exhibit their style and may even be challenging to differentiate from human-produced creations. The broad scope of AI-based language models has led to its applications in diverse fields such as journalism, customer services, health care, market analysis, etc. This paper presents a comprehensive exploration of research databases, official websites, and documents of prominent technology companies to identify the AI-based language models developed in the past five years from 2018 to 2022. It also highlights their impact and usage on understanding and generation of natural language.