G Sunil Reddy
@sru.edu.in
Assistant Professor, Department of Computer Science and Engineering
SR University ( formally S R Engineering College)
RESEARCH INTERESTS
Cyber Security, Computer Forensics, Data Mining, IoT, Network security,
Scopus Publications
Scholar Citations
Scholar h-index
Scholar i10-index
Scopus Publications
G. Sunil, Malatesh Akkur, and Dhiraj Kapila
Springer Nature Singapore
Goli Sunil, Srinivas Aluvala, Chinthala Sujitha, Akarapu Mahesh, Areefa, Kanegonda Ravi Chythanya, and Gadde Aruna
AIP Publishing
Kothandaraman Dhandapani, Rajchandar Kannan, Alavandar Arulmurugan, Goli Sunil, Karuppasamy Kannapiran, Durgunala Ranjith, and Swetha Mucha
AIP Publishing
Goli Sunil, Areefa, Kota Pragathi, Koyyada Rishitha, Sambari Praveen Kumar, Kothandaraman Dhandapani, and Rajasri Reddy
AIP Publishing
Vishali Sivalenka, Srinivas Aluvala, Khaja Mannanuddin, G. Sunil, J. Vedika, and V. Pranathi
AIP Publishing
Sajja Suneel, R. Krishnamoorthy, Anandbabu Gopatoti, Lakshmana Phaneendra Maguluri, Prathyusha Kuncha, and G. Sunil
Springer Science and Business Media LLC
Ricky Rajora, Deepak Banerjee, Mukesh Singh, and G Sunil
IEEE
The classification of performance indicators for each of the ten magnitude scale categories is briefly summarized in the abstract. The model consistently displays an average of roughly 95.25% for precision, recall, or F1-Score metrics across all scales, demonstrating its great accuracy in identifying cases. Support values show that instances are distributed evenly throughout the scales, demonstrating the model's capacity to handle a variety of datasets with effectiveness. Furthermore, the model attains a 99% overall accuracy rate, demonstrating its effectiveness in classifying cases. On a dataset of 14,120 photos, the model achieved an overall accuracy of 85.48 percent. Excellent interpretations are also supported by macro, weighted, and micro average measures of 95.25 percent. In the weighted average, the scores for each class are weighted by its support to counteract class imbalance, while scores for each class are taken into account to give larger classes more importance in the macro average. Micro average, meanwhile, combines all class contributions into a single metric, offering a comprehensive evaluation of the model's effectiveness. Together, these results illustrate how easily the model can spot and categorize examples of different intensities of the same condition. With these consistent results and proven accuracy, a model is a reliable bet for any work that demands precise classification and discrimination. Its potential applications range across a wide variety of fields.
Ricky Rajora, Deepak Banerjee, Mukesh Singh, and G Sunil
IEEE
In the abstract, The article examines a classification model developed to find and classify surface flaws in casting processes, with an emphasis on five unique defect classes: strike, scar, scab, drop, or penetration. The efficacy of the model is thoroughly tested using accuracy, recall, or F1-Score metrics, and support values representing the number of examples in each class. The support proportion highlights the relative prevalence of each class in the sample. The model achieves impressive precision, and recall, with F1-Score values ranging from 97.74% to 98.22% across the various fault classes. The support values vary between 1335 to 1405, indicating the diversity in instances in each class. A precision of 99% demonstrates the model's general ability to make correct predictions. The macro-average, weighted-average, or micro-average metrics provide a comprehensive view of the model's performance, demonstrating its uniformity in predicting defects across categories. The investigation continues by emphasizing the model's durability and possible uses in real-world circumstances, demonstrating its dependability in detecting casting surface defects. The suggested model's high accuracy, as well as recall values, indicate that it can transform inspection processes in the metallurgical industry, resulting in improved product safety, effectiveness, and quality.
Sahil Bhatia, Mansi Sharma, Shweta Pandey, Ruchi Tripathi, G Sunil, and Mansi Sahu
IEEE
Abdul Rahim, Ankur Sharma, Gaurav Sethi, Azad Kumar, G Sunil, and Muntather Almusawi
IEEE
The way we connect to and engage with different things and urban infrastructure has been completely transformed by the Internet of Things (IoT), a paradigm shift in technology. IoT entails providing internet connectivity to various items, such as electrical poles and home appliances, in order to facilitate data collecting and communication. Real-time applications and monitoring are made possible by this connectivity since data is collected and delivered instantaneously or with no delay. The real-time data produced by IoT is crucial in the context of urban development and safety because it provides timely insights into the state of key city assets, ultimately enhancing environmental sustainability and public security. Monitoring the tilt status of electric poles in urban areas in real time is one specialized application of the IoT. To precisely gauge the angle or inclination of these essential infrastructure components, tilt sensing technology is used. The device can determine the stability and safety of electric poles by continuously measuring the tilt angle. An IoT-based solution is suggested to accomplish this, utilizing the NodeMCU development board. NodeMCU is a good option for IoT applications since it easily blends a microcontroller and Wi-Fi module. The NodeMCU serves as the key element in this system and is in charge of gathering data from tilt sensors and sending it over the internet.
S. M. P. Gangadharan, Faz Mohammad, G. Sunil Reddy, Kaushal Kumar, Jagendra Singh, and Haridass R
IEEE
The present research looks at how deep learning models may improve the accuracy of COVID-19 diagnosis using CT scan pictures. The project aims to improve the detection of COVID-19-specific patterns and features by using the capabilities of the VGG19, DenseNet201, and VGG16 models. The training and assessment were based on a collection of 2,521 CT scan images from patients with cardiac problems as well as those who tested positive for COVID-19. To assess the models' diagnostic abilities, performance metrics such as accuracy, precision, recall, and specificity are used. DenseNet201 regularly beats the competition across all key measures, revealing its potential for accurate COVID-19 diagnosis. The suggested edge-cloud architecture further integrates the models into real clinical scenarios. Given the important role of false negatives and false positives in medical diagnostics, the results highlight the need of looking at more than just accuracy. These models are useful diagnostic tools, but they should not be used in place of hands-on medical treatment. This study adds to the increasing body of knowledge at the interface of artificial intelligence and healthcare by providing insights into how deep learning models may be used to improve COVID-19 diagnostic accuracy and perhaps modify medical practices. To fully realize the promise of these findings in enhancing patient care and healthcare systems, more cooperation between machine learning researchers and healthcare practitioners is required.
S. Vinod Kumar, G. Sunil, Ramy Riad Hussein, S. Manju Vidhya, and S. Meenakshi Sundaram
IEEE
Facial Emotion Recognition (FER) plays a crucial role in analysing and understanding individual’s emotions in the form of non-verbal communication. Existing Deep Learning (DL) based FER methods offered diverse contribution in analysing human emotions but several challenges like extracting accurate spatial and temporal features for accurate classification of emotions is a major concern. To address this, a DL based Facial Recognition and Emotion Detection (FRED) system is proposed by using extended Cohn-Kanade (CK+) dataset as the input. The data is pre-processed first where the Histogram Equalization (HE) is applied and then Convolutional Neural Network (CNN) is used to extract the spatial features while reducing the dimensionality. The extracted features are given to Recurrent Neural Network (RNN) to learn long term dependencies from a temporal sequence of data. The final classification is performed where the Attention Mechanism (AM) is integrated in the RNN to determine the higher weights for better outcome. The results of proposed AMConVnet-RNN showed a superior classification accuracy of 99.54%, precision of 98.63%, and f1-score of 97.39% when compared to existing FER methods like Residual Network-50 (ResNet-50) and Channel Based Attention Mechanism (CBAM) based ResNet50.
Sri Lavanya Sajja, Zaid Alsalami, G Sunil, Sunil Kumar V, and Ghazi Mohamad Ramadan
IEEE
Deep Learning (DL) based identification and classification of archaeological sites are being used progressively from satellite images obtained from Geographic Information System (GIS). Even though, there were promising results from the existing researches, there is still need to enhance the classification with distinct features of the satellite images. The methods found challenging to accurately or partially identify an archaeological feature especially in a complex landscape due to lack of distinction between archaeological and non-archaeological features. For this, the geo-referenced satellite images are considered in this research, which are pre-processed using min-max normalization for efficient data analysis and to remove unwanted data. The Selective kernel spatial Feature Extraction (SFE) module is used to improve the feature extraction process where the spatial features are extracted. The extracted features are then classified using Fully Convoluted Deep Neural Network (FCDNN) for effective identification and classification of archaeological sites. The results showed that the proposed SFE-FCDNN achieved better classification accuracy with 98.83%, precision of 97.41%, and an error rate of 1.71 when compared to the existing methods such as Neural Network (NN), and Backpropagation Neural Network (BNN).
Mukesh Soni, G Sunil, N Rajesh, Zaid Alsalami, and Papiya Dutta
IEEE
Wireless Sensor Networks (WSNs) are a substantial part of the Internet of Things (IoT) utilized for cluster and routing path applications to monitor computational fields. The major challenge of the Firefly Algorithm (FA) is its sensitivity to parameters such as the attractiveness coefficient and the randomization factor. Selecting appropriate parameter values can be challenging and may require fine-tuning. The proposed method utilizes energy levels in the IoT to efficiently manage routing data transmission, enhancing information transfer through the cluster router. Aquila Optimization (AO) often involves trade-offs between different parameters. Improving one aspect may negatively impact another, and finding the right balance can be difficult. Although the proposed FA and AO method exhibit high performance compared to existing methods such as Black Widow Optimization (BWO), Fixed Parameter Tractability (FPT), and Multipath Link Routing Protocol (MLRP). The proposed FA and AO method yield high results, achieving a Network Lifetime 1900 rounds, Packet Delivery Ratio (PDR) of 0.98, and an End-To-End Delay (ETED) of 0.80 ms. These outcomes outperform existing methods in the context of WSNs.
Layth Hussein, P Rashmi, G Sunil, Sheeba Armoogum, and S Meenakshi Sundaram
IEEE
Early detection of problems can save wind turbines from significant damage, reducing maintenance costs and downtime. Operators can save millions of dollars in repair and replacement expenses by implementing an efficient fault detection system. Identifying the wind turbine’s issues is challenging for the model to be trained on. This work presents an effective method for detecting wind turbine faults using the Gated Recurrent Unit (GRU) classifier. The SCADA historical wind turbine dataset was first preprocessed using the sliding window approach. Subsequently, the auto-encoder method extracts the spatial features from the processed data. The Multi-Objective Evolutionary Algorithm (MOEA) is then used to select the optimal features. Each individual and the initial population representation consists of fixed-length arrays of bits, with the size matching the number of attributes in the problem that needs to be solved. Finally, the problem is solved using MOEA and the GRU’s effective classifier. The model was evaluated using several metrics, including F1-Score, Accuracy, Precision, and Recall, with numerical values of 1.00, 0.997, 0.805, and 0.999, respectively. A comparative analysis with Convolutional Neural Network-based 2Dimentional Long-Short Term Memory (3DSE-CNN-2DLSTM), Adaptive Neuro-Fuzzy Inference System (ANFIS), and Light Gradient-Boosting Machine (LightGBM) illustrates the current approaches.
Khushi Mittal, Kanwarpartap Singh Gill, Priyanshi Aggarwal, Ramesh Singh Rawat, and G Sunil
IEEE
This research study investigates the use of Deep Convolutional Neural Networks (CNNs) and Transfer Learning in the field of coral reef conservation. It does this by introducing a dataset consisting of 923 photographs that classify corals as either healthy or bleached. The paper aims to solve the labor-intensive process of manually identifying coral reefs by providing an automated picture classification model based on Convolutional Neural Networks (CNNs). This is particularly important due to the crucial function of coral reefs and the issues they face, such as coral bleaching. The approach utilises the VGG 19 pre-trained CNN model and includes the use of training, validation, and testing datasets. Model monitoring is implemented via the use of callbacks. The hyperparameters of the model, such as batch size, epochs, and input shape, are explicitly defined. Common evaluation metrics used in machine learning include accuracy, confusion matrix, and classification report. The study highlights the advantages of the suggested paradigm in terms of efficiency, scalability, and real-time monitoring. The achieved accuracy of 74% demonstrates encouraging outcomes, highlighting the capability of CNNs in promptly identifying and tracking the well-being of coral reefs. This is vital for making well-informed decisions and undertaking conservation initiatives.
Khushi Mittal, Kanwarpartap Singh Gill, Priyanshi Aggarwal, Ramesh Singh Rawat, and G Sunil
IEEE
This study focuses on the crucial matter of identifying and categorising bone fractures via the use of sophisticated image analysis methods. Timely detection of bone fractures is crucial for appropriate treatment due to their prevalence and considerable impact on patient outcomes. Prompt identification is vital to minimise problems and optimise overall recuperation. The research uses a dataset of 4906 X-ray pictures that are categorised into two classes: fractured and not-fractured. It employs a Sequential Convolutional Neural Network (CNN) model. The CNN model achieves an amazing accuracy of 98% in fracture classification using a training set of 4099 pictures, a testing set of 401 photos, and an additional 406 images for validation. The technique highlights the capacity of deep learning in medical imaging to provide precise and efficient fracture diagnosis, demonstrating the potential of this approach to improve healthcare results.
Muskan Agarwal, Priyanshi Aggarwal, G Sunil, Kanwarpartap Singh Gill, and Ramesh Singh Rawat
IEEE
The abstract of this research paper outlines the overarching objectives, methodologies, findings, and implications of the study, focusing on the development of a robust machine learning model for precise classification of various ovarian cancer subtypes. The primary goal was to leverage the VGG19 architecture, a deep convolutional neural network, to accurately identify subtypes including endometrioid carcinoma (EC), clear cell carcinoma (CC), mucinous carcinoma (MC), high-grade serous carcinoma (HGSC), and low-grade serous carcinoma (LGSC). Employing advanced image processing algorithms and feature extraction techniques, the model was trained on a dataset comprising histology images of ovarian tissue samples. The achieved overall accuracy of 75% signifies the effectiveness of the proposed strategy in subtype identification. A comprehensive evaluation of the model’s performance, including accuracy, recall, and F1-score, was conducted to assess its classification skills across diverse subtypes. The study not only presents the model’s achievements but also sheds light on potential areas for growth, future research prospects, and an analysis of the model’s strengths and weaknesses. The research contributes significantly to the field by demonstrating the ability to differentiate between ovarian cancer subgroups, offering valuable insights that could enhance the precision of diagnosis and, consequently, the formulation of tailored treatment regimens. The abstract concludes by emphasizing the increasing trend in the utilization of computational approaches, particularly machine learning, for enhanced accuracy in ovarian cancer diagnosis and treatment. Overall, the abstract serves as a concise and informative preview of the research, inviting readers to delve deeper into the study’s details.
Muskan Agarwal, Priyanshi Aggarwal, G Sunil, Kanwarpartap Singh Gill, and Ramesh Singh Rawat
IEEE
The abstract of this research paper encapsulates the essence of a comprehensive investigation into stellar classification using linear regression. The study explores the predictive power of various features, including Absolute Temperature, Relative Luminosity, Relative Radius, Absolute Magnitude, Star Color, and Spectral Class, with a focus on predicting Star Type. The dataset encompasses a diverse range of stars, from Red Dwarfs to HyperGiants, providing a rich foundation for analysis. The linear regression model demonstrates a remarkable 90% accuracy in predicting star types, underscoring the efficacy of this approach in the context of stellar classification. Precision, recall, and F1-score metrics are employed to evaluate the model’s performance across different star types, revealing insights into its strengths and limitations. The abstract sets the stage for the paper, summarizing the key findings and emphasizing the significance of the research in contributing to our understanding of stars and the broader field of astronomy.
M. Sangeetha, G Sunil, Ammar Hameed Shnain, Shruthi B S, and A U Shabeer Ahamed
IEEE
In building the navigation software application, the Landmark plays an important factor in visual place recognition. There is major problem for performing the classification to visual recognition, due to change of environments like summer, winter, rainy season the place changes. To overcome the problem, the paper proposes a vision transformer as classification for place recognition. In vision transformer the first stage is considerations images are divided into various parts, then it converts into sequence embedding by moving it. To keep information at correct position, the embedding position is used in the parts and the sequence result is sent to various multihead attention layers for final result generation. For validating the proposal, we performed the vision transformer model on the two different datasets such as paris and oxford datasets. The result obtained by performing the vision transformer is demonstrated the other existing model in terms of accuracy in two datasets the given values are 90.63% and 87.30% respectively.
Khushi Mittal, Kanwarpartap Singh Gill, Priyanshi Aggarwal, Ramesh Singh Rawat, and G Sunil
IEEE
This study article focuses on dysarthria, a speech problem that is often seen in patients with cerebral palsy (CP) or amyotrophic lateral sclerosis (ALS). Early identification of dysarthria is crucial for successful treatments and better patient outcomes, since it has a substantial influence on communication ability. Using the TORGO database, which consists of 2000 samples of persons with and without dysarthria, spanning different genders, the research use a Convolutional Neural Network (CNN) to classify speech. The dataset includes both dysarthric and non-dysarthric individuals of both genders, with 500 samples each, captured during separate sessions. The CNN-based technique demonstrates an impressive 96% accuracy in differentiating between dysarthric and non-dysarthric instances, highlighting the promise of deep learning technology in assisting with the early detection of dysarthria and enabling prompt therapies for afflicted people. This study makes significant contributions to the advancement of diagnostic capacities and presents a potential opportunity to improve the quality of life for those with speech difficulties.
Khushi Mittal, Kanwarpartap Singh Gill, Priyanshi Aggarwal, Ramesh Singh Rawat, and G Sunil
IEEE
This research study introduces a novel method for predicting the condition of high-speed bearings in wind turbines. The method utilises unprocessed data collected from sensors that measure vibrations. The process entails calculating a Kurtosis Spectrogram for each of the 50 samples and using a 1D-CNN to categorise bearings into two groups: Medium-life Expectancy (more than 15 days) or Short-life Expectancy (less than 15 days). The research highlights the importance of early identification in forecasting the Remaining-Useful Life (RUL) of machine components, which is vital for averting failures and optimising maintenance choices. The 1D-CNN, implemented using Keras, has remarkable accuracy, achieving an overall accuracy of almost 90% based on recall. This study emphasises the prospective utilisation of this solution in alert systems, offering vital assistance for technical determinations about component upkeep, guaranteeing the dependability and effectiveness of wind turbine systems.
Muskan Agarwal, Kanwarpartap Singh Gill, Priyanshi Aggarwal, Ramesh Singh Rawat, and G Sunil
IEEE
This research introduces a comprehensive approach for enzyme classification, merging Exploratory Data Analysis (EDA) with a Neural Network (NN) model. Leveraging a dataset of 858,777 labeled amino acid sequences, originating from ten organisms, the model predicts enzyme classes in a test set of 253,146 samples while excluding sequences above a set length. Preprocessing involves omitting rare amino acids (X, U, B, Z) and accounting for amino acids (B, Z) exclusive to the training set. The Neural Network architecture, comprising an embedding layer, bidirectional LSTM layers, and a dense output layer, yields promising results with 79% accuracy on the test set after 20 epochs. A detailed classification report and confusion matrix showcase the model's efficacy across 20 enzyme classes. This study advances enzyme classification methodologies, demonstrating the significance of EDA and NN integration in bioinformatics and molecular biology. Future directions involve exploring additional features and optimization strategies for continued model enhancement.
Muskan Agarwal, Kanwarpartap Singh Gill, Priyanshi Aggarwal, Ramesh Singh Rawat, and G Sunil
IEEE
This research study investigates the use of Convolutional Neural Network (CNN) Autoencoders for detecting anomalies in electrocardiogram (ECG) data utilising the PTB Diagnostic ECG Database. The dataset consists of 14,552 samples that are classified into two categories: normal heartbeats and heartbeats influenced by cardiac abnormalities. The use of Transposed Convolution has greatly enhanced the performance of the model. The work highlights the crucial importance of promptly identifying cardiac abnormalities and introduces a CNN Autoencoder model specifically developed to effectively encode and decode ECG data, facilitating the identification of irregular patterns. The process entails constructing a resilient Autoencoder including encoder and decoder components, which are taught to minimise mistakes in reconstruction. The assessment measures demonstrate the model's exceptional accuracy (76.93%), precision (55.23%), recall (89.81%), and F1 score (65.40%). This study emphasises the significance of using deep learning methods to identify anomalies in ECG data at an early stage. This approach shows great potential for enhancing diagnostic skills and improving patient outcomes.
Ricky Rajora, Deepak Banerjee, Deepak Upadhyay, Sarishma Dangi, and G Sunil
IEEE
Their abstract presents a novel method for identifying cassava root disease by creating a sophisticated model. Using Convolutional Neural Networks (CNNs) in combination with Random Forest, the model achieves high recall, high precision, and F1-Score values of more than 97.7% for a few disease classes: cassava Mosaic Disease, cassava Brown Streak Disease, cassava Bacterial Blight, cassava Anthracnose Disease, or cassava Root Rot. The macro-, weighted--, & micro-average accuracy metrics, which add up to an astounding 97.80%, all demonstrate how resilient the model is. With its thorough insights into the model's classification performance, Figure 3 offers a nuanced breakdown of true positives, false positives, false negatives, and true negatives. Furthermore, the complex CNN architecture behind this achievement is explored in Table 3. The research involved examining a dataset of 6810 images. The model demonstrated impressive performance, achieving an accuracy of 97.79% in overall classification. The model's dependability across a range of disease cases is highlighted by these metrics, even though macro, weighted, and micro-average support and support proportion values were missing. The cumulative effect of these results highlights the revolutionary effect of the model on the identification of diseases in cassava production. Enhancing crop health and advancing precision agriculture is made possible by the precision and accuracy of the methodology. The useful tool for farmers to maximize cassava production and secure a more secure and resilient agricultural future, this creative integration of cutting-edge technologies not only transforms disease identification but also makes a significant contribution to the larger goal of sustainable agricultural practices.