International Journal of Intelligent Engineering and Systems

1

Authors: K. Manivannan, S. Sathiamoorthy

Tuberculosis (TB) detection and classification on chest X-ray (CXR) images remains the most significant task in medical diagnosis. TB is a contagious disorder that affects the pulmonary region, and its diagnosing process often depends on CXR. CXR images are utilized for classifying and detecting TB lesions, including infiltrates, cavities, pleural effusions, and nodules. Manual analysis by radiologists includes a visual assessment of the X-ray images by a skilled physician or radiologist. There were many techniques to automatically classify and detect TB on CXR, including deep learning-based approaches, manual interpretation by radiologists, and computer-aided diagnosis (CAD) systems. This manuscript offers the design of pelican optimization with majority voting ensemble model for tuberculosis detection and classification (POMVE-TDC) technique on the CXR images. The core objective of the POMVE-TDC approach is to classify the incidence of TB on the CXR images. At the primary stage, the POMVE-TDC technique undergoes a contrast enhancement process. Besides, the densely connected network (DenseNet-161) model is applied for the extraction of feature vectors. Meanwhile, a pelican optimization algorithm (POA) based hyperparameter optimizer is designed for the DenseNet-161 model. Finally, a majority voting ensemble classifier comprising graph convolution network (GCN), autoencoder (AE), and extreme learning machine (ELM) models are used. The performance evaluation of the POMVE-TDC technique on the medical dataset highlights the significant outcomes with maximum accuracy of 98.83%.

21

Authors: D. Raja, M. Karthikeyan

Plant diseases were responsible for crop losses which directly affect global and national food production methods, ensuing in economic losses. Effectual plant disease analysis includes early-season plant disease detection, detection of several diseases from various crops and various simultaneous diseases, evaluation of the severity of diseases, an estimate of the suitable volume of pesticides for execution, and valuable stages for taking to manage diseases for limiting their spread. Plant diseases and pests were significant features defining crop and plant quality. Plant diseases and pest recognition are applied for digital image processing. Recently, deep learning (DL) is developing enhancement in the field of digital image processing, far higher than standard approaches. This article introduces a red deer optimization with deep learning enabled agricultural plant disease detection and classification (RDODL-APDC) technique. The presented RDODL-APDC technique exploits the DL technique for recognizing and categorizing plant diseases. Initially, the RDODL-APDC technique segments the plant leaf regions using the NestNet background removal process. In addition, multi-level thresholding segmentation segments the infected portions of the images of the plant leaf. For feature extraction, the RDO with the MobileNet-v3 model is exploited. Finally, the extreme gradient boosting (XGBoost) and neural network (NN) models are used for the detection process with particle swarm optimization (PSO) as a parameter tuning technique. For exhibiting the betterment of the RDODL-APDC technique, an extensive range of simulations are executed. The experimental values described the improvement of the RDODL-APDC technique over other current algorithms.

41

Authors: G. S. Sapna, Shashikumar Dandinashivara Revanna

Internet of Things (IoT) is a revolutionary innovation in many aspects of our society like financial activities, communication activities, and global security such as the military and battlefields’ internet. Security and energy play a crucial role in data transmission across IoT and edge networks. In this research, a trust mechanism based on privacy access control is proposed for IoT devices’ interoperability. Most of the existing researches on achieving interoperability for IoT devices has drawbacks such as overlapping of systems, uneven distribution of data, lack of data security, high power consumption, and low optimization of resources. The main objective of this research is to focus and overcome these challenges by introducing a privacy access control mechanism that includes trust parameters of IoT device interoperability. A routing protocol for low-power and lossy networks (RPL) mode of operation is set in the direction of multipoint-to-point traffic flow, except in the downward flow direction. Sensor nodes send data packets to the sink node, which transmits the information to the server to determine the trust values in this mode. To validate the performance, a widely used lightweight low-power wireless simulator Contiki/cooja simulator is implemented. The simulation results of the proposed model have shown a transmission ratio of 100%, a receiver ratio of 30 to 100%, and the detection of malicious nodes in a simulation time of 60 minutes. With the use of the proposed trust mechanism based on privacy access control, a less packet loss ratio of 0.43% is achieved along with less power consumption of 0.4%, and the highest average residual energy of 0.87mJoules at node 30.

79

Authors: Alif Tri Handoyo, Aurélien diot, Hidayaturrahman, Derwin Suhartono, Bart Lamiroy

Sarcasm is a form of figurative speech where the intended meaning of a sentence is different from it literal meaning. Sarcastic expressions tend to confuse automatic NLP approaches in many application domains, making their detection of significant importance. One of the challenges in machine learning approaches to sarcasm detection is the difficulty of acquiring ground-truth annotations. Thus, human-annotated datasets usually contain only a few thousand texts, often being unbalanced. In this paper, we propose two different pipelines of data augmentation to generate more sarcastic data. The first one is SMERT-BERT, a modified SMERTI pipeline that uses RoBERTa as the language model for the text infilling module. The second one is SWORD (semantic text exchange by Word-Attribution), where we modified the masking module in the SMERTI pipeline by utilizing the word-attribution value. These approaches are combined with a SLOR (syntactic log-odds ratio) metric to filter the generated sarcastic data and only select sentences with the best score. Our experiments show that the use of a SLOR filter has a significant positive contribution to the augmentation process. In particular, we achieve the best results when using the SMERT-BERT pipeline and a SLOR filter by improving the F-measure by 4.00% on the iSarcasm dataset, compared to the baseline models.

104

Authors: Sharada Suresh Dambal, Mahesh Kumar Doddananjedevaru, Sudarshan Budnur Gopalakrishna

Heart rate variability (HRV) is the variations in the time interval among heartbeats which indicates the status of the overall cardiac health. Reduced or abnormal HRV is related with an increased death risk. However, the classification of HRV is a challenging task because of the huge number of features processed during the diagnosis. In this research, an effective HRV classification is obtained using the hybrid feature dimension reduction (HFDR) technique along with the support vector machine (SVM) classifier. The HFDR is the combination of kernel principal component analysis (KPCA) and linear discriminant analysis (LDA) which is used to reduce the dimension of the overall feature vector. Three different feature extraction approaches such as continuous wavelet transform (CWT), autoregressive (AR) and spectrogram are used to understand the nonlinear and non-stationary features from the R peak obtained using hybrid R peak (HRP) detection. The reduction in feature dimension using HFDR is used to improve the classification of HRV. The electrocardiogram (ECG) signals from the MIT-BIH arrhythmia database of Physio.Net is used for analyzing the proposed HFDR-SVM. The HFDR-SVM method is evaluated using accuracy, precision, recall, sensitivity, specificity, and F-measure. The existing researches of forward back-propagation neural network (FFBPNN), weight-based ReliefF (WReliefF) with genetic algorithm (GA)-SVM, Alexnet-SVM and long short term memory with angle transform (LSTM-AT) are used for comparing the HFDR-SVM method. The accuracy of the HFDR-SVM is 99.86% that is higher than the FFBPNN, WReliefF-GA-SVM, Alexnet-SVM and LSTM-AT.

124

Authors: Arun kumar Doddanahalli Ramalingegowda, Krishna Alabujanahalli Neelegowda, Bangalore Gnanamurthy Prasad

The recognition of an individual from images/videos captured from different cameras in real-world surveillance scenarios is the Person’s re-identification. We propose image-based person reidentification (Re-id) using GAN, DWT, and MLTP with ANN classifiers. The central server database contains four benchmarked face databases by considering six images per person. The generative adversarial network (GAN) is used to convert side-angled test images into frontal images for an effective person Re-id. The discrete wavelet transform (DWT) and modified local ternary pattern (MLTP) techniques are used on images of the central server database and generated frontal test images to extract initial features. The Final features are computed by using the convolution of the DWT-LL band and MLTP features. The artificial neural network (ANN) is used to categorize and recognize test images. The existing research based on HOG, Hu Invariant moments, CNN, and SVM is presented and published, having a recognition accuracy of 97%. However, our projected technique uses GAN, which improved recognition accuracy to 99% compared to existing techniques having low recognition accuracy.

149

Authors: Lucky Indra Kesuma, Ermatita, Erwin

A Chest X-ray (CXR) image can diagnose lung diseases. However, a diagnosis requires time and high accuracy, so an automatic system is needed. Convolutional neural network (CNN) is a reliable method for image classification and has many architectures. ResNet is a CNN architecture that can overcome gradient vanishing, but it has a deep network structure to detect errors. The EfficientNet CNN architecture can proportionally uniformize all depth, width, and resolution dimensions in each layer as needed, but it takes a long time in training. The Inception-v3 CNN architecture uses Inception blocks by reducing dimensions to small convolutions, but it has larger parameters than other architectures. ELREI is an acronym for ensemble learning of ResNet, EfficientNet, and Inception-v3 with weighted voting. ELREI combines the classification results on the ResNet, EfficientNet, and Inception-v3 architectures to overcome the limitations and combine the advantages of each architecture. ELREI works on the training stage at each epoch rather than the final results of each architecture. In addition to voting, ELREI uses a fully convolutional Network (FCN) at the final stage for the best weight determination and to prevent overfitting during training. The results of the accuracy, precision, recall, and F1-score of the ELREI method are excellent, above 98%. The training graph of the ELREI ensemble method proves that ELREI can overcome overfitting that occurs on the architectures. The results show the Ensemble ELREI method is excellent and robust for lung disease classification based on CXR images, which are carried out in 4 classes: normal, COVID-19, lung opacity, and pneumonia.

173

Authors: Shrikanth Shirakol, S S Kerur

The Discrete Cosine Transform (DCT) is a basic transform block used in Adaptive Multicore Transform (AMT), which is a core of High Efficiency Video Coding (HEVC) and Versatile Video Coding (VVC) standards. AMT uses artificial intelligence technique to decide on the transform output. The suggested One-Dimensional DCT architecture is conceived by 8-point structure Loeffler-DCT technique and synthesized using a floating-point Arithmetic. By preserving the structural regularity as the Loeffler-based design, the optimized floating-point architecture consumes optimum space and delay thereby increasing the precision. The work focuses on obtaining high precision output without compromising on ADP (Area-Delay-Product). Due to the fact that the floating-point multiplier unit is developed using shift-add operations, the results reveal the accomplishment of better resolution output while maintaining Root Mean Square deviation as low as 0.0062. A total of 117 additions, 66 shifts are employed in the suggested architecture. The proposed 1D-DCT is used as a sub-block in developing the architecture of 2D-DCT using only 50% of the 1D-DCT subblocks that is needed for the conventional technique. The 8 X 8 2-D DCT, is computed using only 8 1-D DCT's and additions, instead of using 16 1-D DCT's, as in the traditional row-column method. The model is tested with standard images, which resulted in better PSNR and MSE compared to the standard method. In comparison to the state-of-the-art on DCT, the proposed method obtains a root mean square error that is negligible up to three decimal places resulting in improvement of PSNR by 17%, with a maximum clock frequency of 496MHz has been achieved. The proposed design strikes a balance between trade-off parameters such as area, speed, and precision.

198

Authors: Yathish Aradhya Bandur Chandrashekariah, Dinesha Hagare Annappaiah

Big data refers to the large amount of information that is collected from different areas and shared on the internet. However, this development has led to difficulties in using frequent itemset mining applications. To overcome the issue of frequent data mining, this research has introduced an empirical sampling algorithm using RadeMacher average (ESA-RMA). When considering the size of the initial sample and scheduling the samples, the ESA utilizes the RadeMacher average to bound the samples. Initially, the data is obtained from the dataset of the human activity recognition (HAR) and real time datasets from smartphone gyroscope and accelerometer, then obtained data is pre-processed using the data normalization technique. Then, the ESA is used to select the labelled data and RMA is used to bound the samples. This bounding process defines the upper limit of the input data which helps in the effective mining of frequent item sets. Thus, the data with redundant items are mined out using the proposed ESA-RMA method. The experimental results show that the proposed ESA-RMA has taken a minimum run time of 212 ms for data obtained from smartphone accelerometer which is comparatively lower than the existing Scalable Simple Random Sampling (ScaSRS) with processing time of 362 ms. Similarly, for HAR dataset, the proposed method took processing time of 5.43 s whereas the existing vertical frequent time interval–related pattern (VertTIRP) mining approach took processing time of 7.82 s.

226

Authors: Mochammad Sahal, Trihastuti Agustinah, Achmad Jazidie, Haiping Du

Agent-based moving source-seeking problem with strategic control considering obstacle avoidance is addressed in this paper. The agents communicate with one another using a communication topology to coordinate finding the object. The source or seeking target is represented by a scalar field, whereas the presence of the obstacle is represented by an artificial potential field. A moving source that moves linearly from the initial point to the end point is considered. Several obstacles, which needs to be avoided by agents, that may be stationary or moving are considered. Double integrator agents are considered, which contains richer dynamics compared to the single integrator agents. The strategic control law which is a combination of gradient consensus and distributed velocity control affected by artificial potential field are used to determine the velocity of each agent to track the moving source while avoiding the obstacles. Formation control is used to make sure that the agents stay under the desired formation for most of the time. The computer simulation is used to examine how different agents would search for a moving source under stationary and moving obstacles. In both cases, the agents can find the source with a different seeking time: 62.8s for stationary and 90s for moving obstacles. Comparing with virtual artificial potential field method, our proposed method manages to perform better in the case of shrinking agents for stationary obstacles.

251

Authors: Gowtham Rajendiran, Jebakumar Rethnaraj

Aeroponics is a popular soilless crop cultivation technology that integrates plant nutrition, physiology, and ecological control. It offers automated monitoring, protected cultivation, improved growth mechanisms, better yield and requires less maintenance. Here, to predict the crop yield, two systems are available: manual and automated. Manual systems often fail to produce better prediction results, leading to substantial crop losses whereas, the automated systems use machine intelligence for growth monitoring. This article proposes a lettuce crop growth monitoring-boost (LCGM-Boost) regression model for lettuce yield forecasting in aeroponic vertical farming system. This model is highly robust to outliers, produces better prediction results of 95.86% and lower error rates of 0.36 (MAE), 0.40 (MSE), and 0.63 (RMSE) than other machine learning models namely, support vector, random forest and XGBoost regressors. Hence, it is preferable for growth monitoring and yield prediction of the lettuce crop in the real-time aeroponics system.

275

Authors: Purba Daru Kusuma, Anggunmeka Luhur Prasasti

This work offers a new stochastic optimization i.e., a metaheuristic algorithm combining both direction-based search and neighbourhood search called as walk-spread algorithm (WSA). These two types of searches become the inspiration for its name where the term walk represents the direction-based search while the term spread represents the neighbourhood search. There are two direction-based searches performed in every iteration where each search produces a single child. Meanwhile, there are two neighbourhood searches performed in every iteration where each search produces several children. The global best unit becomes the first reference while two shuffled units become the second reference in performing the direction-based search. Meanwhile, the local search space of the first neighbourhood search is wide while the second one is narrow. The 23 classic functions are chosen as the assessment of WSA where WSA is confronted with the five latest metaheuristics: mixed leader-based optimization (MLBO), golden search optimization (GSO), pelican optimization algorithm (POA), zebra optimization algorithm (ZOA), and attack-leave optimization (ALO). The assessment result shows that the offered WSA achieves the acceptable result so fast. Moreover, WSA is also superior to these five confronters by outperforming MLBO, GSO, POA, ZOA, and ALO in 23, 23, 22, 21, and 21 functions respectively.

301

Authors: M. Arunadevi, V. Sathya

The on-demand provision of computing resources as services over the internet is known as cloud computing. The distributed denial of service (DDoS) attack is a major security risk that affects cloud services. Because of the computational complexity that must be handled, detecting DDoS attacks is a very difficult operation for cloud computing. The back propagation neural network (BPNN) method is frequently employed for DDoS attack detection due to its great flexibility and straightforward construction. But it has drawbacks such as slow convergence, inconsistencies, and instability during training. In this work, the proposed optimized BPNN uses bacterial colony optimization (BCO) for optimizing the connection weights and biases to enhance the performance of BPNN. The optimized BPNN is developed to identify DDoS attacks in the cloud environment. The performance of the BCO-BPNN detection scheme is assessed using four DDoS attack datasets such as NSL-KDD, ISCXIDS2012, CIC-IDS2017, and UNSW-NB15. Its respective detection accuracy with the NSL-KDD, ISCXIDS2012, CIC-IDS2017, and UNSW-NB15 datasets is 0.9892%, 0.9883%, 0.9341%, and 0.9987%. The results of the studies demonstrate that the suggested BCO-BPNN performs better than existing BPNN variants, conventional BPNN, and support vector machine (SVM) methods.

329

Authors: Chandini Shivaramu Banumathi, Ajjipura Basavegowda Rajendra

In recent years, the growth of malicious software is becoming high and led to a huge number of security threats. The malware exploits the system information and uses the important information of the user without any intimation. Moreover, the malware furtively directs that information to the servers which are organized by the attackers. In recent years, researchers and scientists discovered anti-malware products to identify known malware. But, these methods are not robust to detect obfuscated and packed malware. The fore mentioned issues rely on the existing approaches can be rectified using the hybridized least absolute shrinkage selection operator and cross validation (LASSO-CV) algorithm are proposed. The LASSO is used in the process of selecting the stochastic features and CV visualize the feature using weighted co-efficient approach. The effective features are selected by the LASSO-CV algorithm which eases the process of classifying the malware. This research also utilized long-short term memory (LSTM)-softsign classifier to classify the malware. The malware samples are collected from the VXHeavens and android malware dataset (AMD) dataset which consists of malware samples from various software. The classification results obtained from the LSTM-softsign classifier have better classification accuracy of 97.85% for VX heavens dataset and 99.24% for AMD.

350

Authors: Huda W. Al-Dulaimi, Ahmed Aldhahab, Hanaa M. Al Abboodi

The acoustic features extracted from the speech-signal are a critical challenge for implementing an accurate speaker identification system. In this paper, two-dimension discrete multi-wavelet transform (2D-DMWT) in conjunction with the deep learning neural networks are proposed for speaker identification. The DMWT is based on a vital sampling scheme preprocessing that uses the filter invented by Geronimo, Hardian, and Massopust, which is call GHM. The system proposed involves firstly preprocessing in which the speech-signal is resampled into 16kHz. Then, the speech-signal is divided to five different durations: 0.5 sec., 1 sec., 2 sec., 3 sec., and 5 sec. In this paper, each duration is tested separately. Second, 2D-DMWT is employed to obtain discriminant features from the speech-signal and reduce speech-signal dimensions in the feature’s selection phase. Finally, neural network algorithm based on convolution neural network (CNN) is used for classification. The system proposed is tested using four databases: SALU-AC, ELSDSR, TIMIT, and RAVDESS. These databases include various speech variances, such as age, gender, etc. The results obtained by the proposed system are 95.86%, 96.59%, 89.90%, and 89.83% for 0.5sec of the SALU-AC, ELSDSR, RAVDESS, and TIMIT databases, respectively. For 1sec, the SALU-AC, ELSDSR, RAVDESS, and TIMIT databases obtained 96.30%, 97.31%, 96.05%, and 93.59%, respectively. The SALU-AC, ELSDSR, RAVDESS, and TIMIT databases achieved 96.63%, 97.76%, 96.12%, and 95.90%, respectively, over the 2sec time duration. During the time duration of 3sec, the SALU-AC, ELSDSR, and RAVDESS databases obtained 97.04%, 98%, and 97.96%, respectively. For 5sec, the SALU-AC, and ELSDSR databases attained 97.56%, and 98.30%, respectively. The results accomplished by the proposed system are outperformed those results discussed in the previous works based on the same databases.

376

Authors: Ameer Ali, Saad Hreshee

Generalized frequency division multiplexing (GFDM) presents an outstanding performance in out-of-band, latency, and peak-to-average power ratio, Thus, promoting it to be the most contender for the fifth generation and beyond. This paper evaluates the performance of the GFDM transceiver system with polar code (PC) and interleavers based on the bit error rate (BER). The proposed system is enhanced by artificial intelligence, which is the genetic algorithm (GA) used to optimize the parameters of the pulse shaping filter (G) and the artificial neural network (ANN) used in improving method-based least square (LS) channel estimation dataset. At each step of evaluating the performance computed based on BER vs. SNR. The optimized PC-GFDM enhances over standard GFDM by 0.0649 in BER at 28 SNR.

395

Authors: Ali Hamzah Najim, Ali Hashim Abbas, Kareem Al-sharhanee, Hussein Muhi Hariz

Internet of drones (IoD) are commonly constructed with unmanned vehicles, have been progressively prevalent due to their capability to operate quickly and their vast range of applications in a variety of real-world circumstances. These IoDs are interact with zone service providers (ZSPs) to achieve the goal of assisting drones in accessing controlled agriculture services. The utilization of drones in precision farming has lately gained a lot of attention from the scientific community. This study addresses with the assistance of drones in the precision agricultural area by analysing communication protocols and applying them to the challenge of commanding a fleet of drones to protect crops from parasite infestations. objectively and equitably assigns a weight to multiple service scheduling parameters based on maldistributed decision making theory, calculates the serving priority of each service request group, and then serves the service request groups based on the calculated serving priori-ty accordingly Hence, this paper proposes reinforcement learning-based topology-aware routing protocol with priority scheduling (RLTARP) to provide reliable combinations between the source and destination. It also improves the routing decision by considering two-hop neighbour nodes, extending the local view of the network topology. The priority scheduling method adopts maldistributed decision making theory, to find the group of priorities based on service request. The proposed RLTARP is compared with three existing methods such as DroneCOCoNet, Markov decision process (MDP)and deep deterministic policy gradient (DDPG) and hence it produces 46.34% of packet delivery ratio, 67.49% of end to end delay, 16.45% of routing overhead, 13% of energy consumption and 97.6% of network lifetime.

417

Authors: Misbah, Muhammad Rivai, Fredy Kurniawan

Diabetes is a disease that the entire world is afraid of. Early detection measures are required in this instance. Diabetes diagnosis by blood samples has the disadvantage of being uncomfortable. This study uses a urine sample to estimate a person's risk of developing diabetes. The electronic nose is a system that has the potential to recognize diabetes in this way. This method consists of quartz crystal microbalance gas sensors coated with carbon nanomaterials, including single-walled carbon nanotubes, double-walled carbon nanotubes, multi-walled carbon nanotubes, and graphene oxide. A reciprocal counter implemented in a field programmable gate array (FPGA) device is used to measure the frequency shift on the sensor. The convolutional neural network technique is used to detect diabetes. The results of the experiments suggest that this system can distinguish between healthy and diabetic people with an accuracy of 91%.

438

Authors: Saja Abdul-hamza, Laith A. Abdul-Rahaim, Sarmad Ibrahim

Electricity theft is a widespread problem with significant economic implications. In this paper, we present an efficient system to detect electricity theft using cloud computing and deep learning techniques, specifically convolutional neural networks (CNNs) due to its ability to effectively extract features from electrical data. To address the challenge of imbalanced data, we tested various data preprocessing techniques. The adaptive synthetic (ADASYN) technique was selected to handle data imbalance. The evaluation of our CNN model demonstrates its effectiveness in accurately detecting incidents of theft, where our model achieved an accuracy of 97.22%, with a precision of 97% and a recall of 99.9%. The system was tested under real-life conditions and proved to be effective. Furthermore, cloud computing techniques have greatly facilitated the dissemination of CNN models by providing storage for data and a computational space for executing the model, as well as presenting the results to the authorities for electricity theft detection. The use of cloud servers has greatly simplified the distribution and utilization of CNN models in the context of electricity theft detection.

462

Authors: M. J. Al-Doori, M. F. Al-Gailani

In this paper, we have presented a comprehensive study optimizing the performance of the N-th degree truncated polynomial ring units (NTRU) post-quantum cryptography (PQC) on advanced RISC machines (ARM)-based internet of things (IoT) devices, with specific focus on Raspberry Pi boards, for integration within edge layer and fog-based IoT architecture. Our objective was to expedite encryption and decryption processing times for five security levels. To achieve this, we adopted Shamir's Secret Sharing (SSS) to facilitate secure NTRU public key exchanges and utilized run-length encoding (RLE) to compress ciphertext effectively. Our results confirmed a compression ratio of approximately 69.25%, suggesting a strong potential for efficient data transfer. The empirical outcomes of our research evidenced considerable improvements in NTRU performance on ARM-based devices, with encryption and decryption times improvement by 80% and 33%, respectively, for 128-bit payloads.

488

Authors: Yunifa Miftachul Arif, Hani Nurhayati, Ahmad Fahmi Karami, Fresy Nugroho, Fachrul Kurniawan, Harits Ar Rasyid, Qurrotul Aini, Norizan Mat Diah, Manuel B. Garcia

Serious game is one of the pedagogical media capable of transferring knowledge to its players. This game genre requires a support system that adaptively selects the appropriate scenario for players to increase their interest and comfort. Therefore, this study proposed an adaptive scenario selection (ASS) system using a finite state machine based on an artificial neural network (ANN). The game scenario is selected by ASS based on five player preferences, including work, hobbies/interests, origin, group members, and repetition. Furthermore, the multi-layer perceptron (MLP) architecture was used in the scenario selection process for the proposed ANN method. The experimental stage was carried out using the theme of travel in several tourism destinations in Batu City, East Java, Indonesia. The experimental results show that ASS succeeded in generating adaptive game scenario choices for players based on their preference data with an accuracy of 67.25%.

512

Authors: Rajalekshmy Kayiparambil Devarajan, Sheeja Shaik Khader

In medical analysis, early detection and diagnosis of cerebral palsy (CP) are highly helpful in stimulating brain cells and alleviating the adverse impacts of the disease. In the detection of CP, diagnostic tools such as magnetic resonance imaging (MRI) and general movements assessment (GMA) have emerged. In the past years, most studies have focused on GMA-based CP detection based on the different pose estimation methods. Amongst, the part affinity field (PAF) or OpenPose was one of the well-known methods for estimating the skeletal images from the RGB-D videos of infant general movements, which was used for CP detection. But, the estimated skeletal images were very few and it was difficult to annotate a large infant movement dataset. Therefore, in this article, a pose sequence-aware generative adversarial network (PS-GAN)-based data augmentation method is proposed that creates high-quality skeleton images for CP detection. First, long-range dependencies in continuous frames are acquired by self-attention and the dense graph is pruned to achieve efficient training. Then, spatial joints and temporal characteristics are encoded into the PS-GAN using the graph convolutional network (GCN) to map noises to high-quality skeleton images. Besides, the PS-GAN structure selection problem is defined as a Markov decision process (MDP) and solved using the new reinforcement learning (RL) to choose the optimal PS-GAN structure, which achieves effective generation. Further, the created skeleton images are used to train the convolutional neural network (CNN) with a softmax classifier and detect CP. At last, an extensive experiment is conducted on the MINI-RGBD, babyPose and motion infant analysis (MIA) databases. The results show that the PS-GAN-CNN achieves 92.2%, 92.5% and 92% accuracy for MINI-RGBD, babyPose and MIA databases in contrast with the PredictMed, fully connected network (FCNet), CNN-long short-term memory (LSTM) and knowledge-based recurrent neural network (KBRNN) methods.

539

Authors: M. Sumithra, N. Rajkumar

Recognizing the facial expressions from the given input is one of the challenging and demanding tasks in recent days owing to the low-resolution images and different backgrounds. Also, the facial emotional/expression recognition system has gained a significant attention in the field of computer vision. The conventional works implemented a variety of machine learning algorithms for face emotion recognition, but higher computation costs, decreased reliability, redundant information, and increased computational time requirement are still only a few of the issues. In this case, an image face filtering method is employed in the beginning to produce a normalized output image with high contrast and quality, which is used to increase the classifier's overall rate of emotion recognition. Then, the novel pelican optimization algorithm (POA) is used to optimize the feature set in order to guarantee the success of classifier training and testing. Prior to emotion recognition, this technique is used to reduce the dimensionality of information. Based on the given face image's optimized attributes, the weight optimized extreme learning machine (WOELM) algorithm is used to reliably identify the emotions. By using well-known datasets like CK+ and MMI, a wide range of factors are taken into consideration for study in order to compare and evaluate the outcomes of the proposed POA-WOELM face emotion identification system. The obtained results reveal that the combination of POA-WOELM provides an increased 99% of emotion recognition accuracy to all kinds of emotions, with the sensitivity of 98.8% and specificity of 98.9%.

562

Authors: Mahantesh H M, Nageswara Guptha M, Hema M S

SDN (software defined network) is a networking approach which enables the IoT (internet of things) to be maintained through programming. The assessment of best routing path satisfying the requirements of QoS (quality of service) possess a significant role in networking. Several methods have been employed by the researchers for maintaining the necessities of QoS. But due to certain limitations like data latency, protocol complexity and bursty traffic which affects the network performance. In the proposed work, adaptive sparse bayesian-find fix finish exploit analyze- extreme learning machine (ASB-F3EA-ELM) is employed with four steps such as cluster formation, cluster head selection, rule caching policy and optimal path selection. The cluster formation is performed by using grid structures for forwarding the data packets. To avoid data transfer from all the IoT nodes, process of cluster head selection for each cluster is performed by modified aquila optimizer (MAO). A prominent rule caching policy is implemented for reducing rule caching cost in SDN based IoT architecture. The best routing path for effective data transfer is performed by the approach of ASB-F3EA-ELM. The performances of proposed approach are analyzed in the NS2 platform. The outcomes are estimated in terms of bandwidth, rule caching cost, delay, throughput, network lifetime and packet loss rate. When comparing to the existing methods, the proposed work shows better output in satisfying the requirements of QoS as it achieved improvement in throughput of 3.89% to 7.97%, packet loss rate reduced from 0.11% to 0.07%, network lifetime increased by 13.75% to 27.5% as compared to PSO, GA, ACO and GWO methods.

586

Authors: Hadeel H. Alfartosy, Hussein K. Khafaji

Text classification is the automated technique used to classify text into a predefined category that is more related to text. Most studies and researches have focused on classification texts written in English rather than Arabic because of the Arabic nature and difficulty of its structures. The difficult nature of Arabic makes it more complex and difficult to deal with because of its many rules and characteristics that are unique to it, but it has become necessary to deal with this language because of its widespread use on the internet. Feature representation and extraction, especially for text, have attracted considerable attention in recent years. The main object of this process is to convert text into a numerical representation. All approaches and systems depend on the frequency of words within text, but these approaches are few. This paper presents a new feature extraction method aimed at furthering natural language processing applications in any language, especially Arabic. It is based on transforming a vector of representation in the time domain into the frequency domain. Fourier feature extraction is a powerful and versatile technique. The primary goal of this approach is to extract salient features from raw text, and then a filter is used to remove noise from the extracted features. Its scalability and efficiency make it suitable to be used on large datasets, making it a widely adopted feature extraction technique. For classification, we used logistic regression, which is a powerful tool for classifying text data and offers significant advantages such as speed and accuracy. We used three world datasets (CNN, OSAC, and SANAD), and the results showed a good performance. The proposed method outperformed recent approaches, in which accuracy reached 98%.

612

Authors: Yahya Bougteb, Elyazid Akachar, Brahim Ouhbi, Bouchra Frikh, Elmoukhtar Zemmouri

Recommender systems are extensively employed across various domains, encompassing movies, hotels, and restaurants, with the aim of delivering personalized recommendations to users. However, the conventional approach of relying solely on a single rating may fall short in comprehending the underlying factors contributing to a user's overall rating. As a remedy, multi-criteria recommender systems (MCRSs) have emerged as an alternative paradigm, allowing users to rate items based on multiple dimensions. The primary challenge lies in predicting the overall rating considering a user's multi-criteria scores. In this study, we propose a MCRS that leverages a multi-criteria decision-making method, which takes into account the interdependence among criteria. To this end, the weight assigned to each criterion is calculated using the correlation coefficient and standard deviation (CCSD) approach. Furthermore, we employ a deep autoencoder to capture intricate nonlinear relationships and generate recommendations of heightened accuracy. Evaluating our proposed method on the well-known MovieLens and TripAdvisor datasets, we demonstrate its superior performance compared to several single and multicriteria recommender systems. The results highlight the effectiveness of our approach in capturing pertinent aspects and achieving lower prediction errors, showcasing its potential in enhancing recommendation quality and personalization.

637

Authors: Aulia Teaku Nururrahmah, Tohari Ahmad

The intrusion detection system (IDS) has been developed to detect attacks or suspicious activity on a network. IDS are generally classified into two types: signature-based and anomaly detection. Many studies widely use anomaly-based detection because it can detect new types of attacks on the internet network, but it has several shortcomings. Handling data with high dimensionality directly to the classification process could lead to low accuracy and increased false alarm rates. Selecting relevant features and removing irrelevant features from classification results could be the solution to overcome this issue. In this research, we propose an intrusion detection model using a combination of the Chi-square independence test and an exhaustive search. Firstly, this proposed method employs the independence levels of the Chi-square test to calculate the statistical scores for each feature. The feature list obtained from the first process is continued to the optimisation stage using an exhaustive search. This process aims to calculate the accuracy values of all possible feature combinations from the early-stage feature list and check each feature combination to see if that combination has the best accuracy. This method was tested on four datasets: KDD Cup 99, NSL-KDD, Kyoto 2006+, and UNSW-NB15 using three classifiers: Support vector machine, decision tree, and Naive Bayes. This method achieved the highest accuracy when tested on the UNSW-NB15 dataset using the SVM. Accuracy, precision, recall, and F-score reached values above 95%. Likewise, the FPR value reached the lowest rate of 1.56%.

666

Authors: Kamadenahalli Narayanaswamy Prema, Mandalur Basavarajappa Ushadevi

Wireless mesh networks (WMNs) are familiar due to their unique characteristics such as adaptability, flexibility and less time for transmitting data packets. The routing technique acts a significant role in transmitting data among the nodes. Moreover, the routing technique is mainly responsible to improve the performance of WMNs with various applications. With the rapid development of various applications, energy, and distance are fundamental aspects which aid in next-level communication. But, recent research has not provided a clear idea of improving the transmission quality by offering an optimal route. These challenges are overwhelmed using the proposed multi-objective levy flight artificial rabbit optimization algorithm (ML-AROA) for energy and distance aware (EDA) route discovery. The proposed ML-AROA is an improvisation of the ARO algorithm where the Levy flight technique is implemented to create a random number in a regular manner which is characterized by the generation of arbitrary numbers and helps to jump out the local solutions. Moreover, the convergence accuracy gets enhanced by offering flexibility to detect an optimal route for the transmission of data packets. The results obtained from the experiment shows that the suggested ML-AROA attained better network throughput of Kbps which is comparatively higher than load balance and interference avoid-partially overlapped channels assignment (LBIA-POCA) framework and multi-objective dyna Q based routing (MODQR) technique.

688

Authors: Tuba Batool, Rozaida Ghazali, Irfan Javid, Nureize Binti Arbaiy

Solid waste management is an imperative aspect of a hygienic environment. Lack of attention towards solid waste management leads to environmental pollution which causes catastrophic health issues and reduces the quality of life. Therefore, it’s required to develop environmentally safe and adequate protocols to manage solid waste for the fortification of human health and the environment. The study presents a master plan to facilitate the local municipal authorities in developing a better waste management framework. The main aim of the research is to predict municipal solid waste (MSW) collection by using the GRU model in Multan city, Pakistan. The GRU model has utilized a dataset collected through a weighbridge at Habiba Sial landfill site (HSLS) which is the most reliable source for the data. The dataset consists of monthly data for the past five years with a total of 180 records; 60 records for months 60 records for vehicle trip numbers and 60 for collected MSW. The number of months and monthly vehicle trip numbers have been used as the input data and monthly collected MSW has been used as output data. The window of time for collected data is June 2017 through June 2022. The dataset has been divided into 3 subparts where 70% of the data has been utilized for model training, 15% of the data has been used for model testing and 15% of the data has been utilized for validation purposes. The GRU model’s performance has been evaluated by using the performance metrics; mean square error (MSE), and Regression values for training, testing, and validation data. According to the results, 2-5-2-1 topology with 5 hidden neurons gives the best results with the least value of MSE being 0.0388 and the maximum value of regression being 0.94. The study also proposed a master plan for waste management, considering options like landfilling, recycling, incineration, and composting. As the landfilling method is the most common practice to dispose of waste, therefore the study estimated the required landfill area based on the predicted solid waste generation rate and its collection rate by the year 2031 for Multan city, Pakistan. The master plan can be utilized alternatively, especially for the cities with the same type of demographics which may help the municipal authorities for the development of a better solid waste management framework.

717

Authors: Suphian Mohammed Tariq, Intisar Shadeed Al-Mejibli

Determining the localization and tracking of sensor nodes in indoor environments is the goal of this study. The difficulty of significant estimation errors in target localization brought on by unpredictable noise in received signal strength indicator (RSSI) readings, particularly in indoor environments, is a major area of research right now. This study proposed a hybrid technique called particle swarm optimization- generalized regression neural network (PSO-GRNN) to improve the sensor nodes' ability to estimate location and target tracking with more precision, as an alternative to conventional RSSI-based strategy. The GRNN algorithm can use the RSSI values as initiation data of the algorithm of GRNN to locate the location and tracing of the target node. An essential component of the GRNN architecture is the spread constant(σ), The method of trial-and-error to select a value of spread constant(σ), which is insecure and does not always yield the best result, is used to choose the parameter. The PSO method is used to determine the optimal GRNN spread constant value. The hybrid PSO-GRNN method was used to overcome these drawbacks and enhance localization and target tracking accuracy without the need for further apparatuses. The tracking algorithm PSO-GRNN the hybrid outperformed the conventional LNSM technique and produced impressive results. Comparing the proposed method to the conventional RSSI, a considerable gain of 87.58% is possible.

738

Authors: Malathi Salejangal Yallappa, Geeta Rudragouda Bharamagoudar

Knee osteoarthritis (OA) is a prevalent condition that typically affects elderly individuals. It occurs due to the gradual deterioration of hyaline cartilage located between knee joints. If it not treated in earlier stage, it causes knee replacement so the knee OA early diagnosis is essential for better treatment. The knee OA diagnosis includes knee X-ray images and classifying them using kellgren-lawrence (KL) grading system. This paper proposed the application of pre-trained models named as skip connection based ResNet101 for knee OA images from osteoarthritis initiative (OAI) dataset. The skip connection based ResNet101 is utilized to overcome the vanishing gradient problems. Moreover, this skip connection allows learning of the recognized functions. This enables the ResNet101 model have performance high-level layers compared to the low-level layers. Due to these features the skip connection based ResNet101 model is utilized in this paper. The two various classifications are performed in this paper such as binary and severity. The binary classification determines whether knee osteoarthritis (OA) is absent or present, which also categorizing the severity of KOA into three grades. The experiments are conducted on three various datasets named dataset I, II & III with five, two and three grades of knee OA images. The obtained result shows that the skip connection based ResNet101 model achieves better accuracy of 79.94%, 84.96% and 95.86% in the above-mentioned three datasets when compared to ResNet101 and InceptionNetV2 models.