Pugalendhi Ganesh Kumar
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Pugalendhi Ganesh Kumar has a total of 72 co-authors. Here are the top co-authors by number of shared publications:
- T. Aruldoss Albert Victoire
Department of Electrical and Electronics Engineering, Anna University Regional Campus, Coimbatore, India
6 shared publications D. Devaraj
Dept. of Computer Science and Engineering, Kalasalingam University, Krishnankoil
5 shared publications- Durgadevi Velusamy
Department of Information Technology, Sri Sivasubramaniya Nadar College of Engineering, Kalavakkam 603110, India
5 shared publications - Arul Antran Vijay Subramanian
Karpagam College of Engineering, Coimbatore, India
4 shared publications C. Rani
4 shared publicationsKu-Jin Kim
Anna University Regional Campus
4 shared publications
Latest Publications
- Open Access
- 11 November 2026
- Published by Frontiers Media SA in Frontiers in Environmental Science
- Vol. 14, 1894362
An IoT-enabled environmental informatics framework for real-time crop disease detection using attention-driven deep learning
Introduction: Crop diseases are a major challenge in agriculture because their early and accurate identification is essential for reducing crop losses and improving productivity. The integration of the Internet of Things (IoT) with deep learning provides an effective approach for automated crop disease monitoring by enabling the collection and analysis of environmental parameters such as soil moisture, temperature, and humidity. However, accurate fine-grained disease identification remains challenging because of variations in disease symptoms, environmental conditions, and visual characteristics. Methods: This study proposes an IoT-enabled deep learning framework for automated crop disease detection and classification. The collected crop images are pre-processed using scaling, normalization, and data augmentation techniques to improve image quality and model generalization. For efficient disease detection, a lightweight attention-driven convolutional neural network-based object detection approach, termed YOLO10-YSTs (You Only Look Once 10-Yield Surveillance Technology), is employed. The framework integrates the real-time detection capability of YOLO10 with an attention mechanism to emphasize disease-relevant regions and a stigmergy-based encoding strategy to enhance feature representation and extraction. In addition, a novel DBGU-LSTM (Dual-Bi-Gated Recurrent Unit with Long Short-Term Memory) architecture is developed for fine-grained crop disease classification. Results: The experimental results demonstrate that the proposed framework provides effective disease detection, discriminative feature representation, and fine-grained classification. The DBGU-LSTM architecture achieves improved classification performance and training stability compared with the conventional deep learning models considered in the study. Overall, the proposed approach achieves a classification accuracy of 97.99%, demonstrating its effectiveness for automated crop disease recognition. Discussion and conclusion: The results indicate that combining IoT-based environmental monitoring with attention-driven deep learning and recurrent feature-learning mechanisms can provide an effective solution for intelligent crop disease detection. The proposed framework enables accurate identification of disease patterns while supporting efficient and automated agricultural monitoring. Therefore, the approach has potential to assist farmers in early disease diagnosis, reduce crop losses, and improve agricultural productivity.
- 11 November 2026
- Published by Frontiers Media SA in Frontiers in Environmental Science
- Vol. 14, 1894362
Open AccessAn IoT-enabled environmental informatics framework for real-time crop disease detection using attention-driven deep learning
Introduction: Crop diseases are a major challenge in agriculture because their early and accurate identification is essential for reducing crop losses and improving productivity. The integration of the Internet of Things (IoT) with deep learning provides an effective approach for automated crop disease monitoring by enabling the collection and analysis of environmental parameters such as soil moisture, temperature, and humidity. However, accurate fine-grained disease identification remains challenging because of variations in disease symptoms, environmental conditions, and visual characteristics. Methods: This study proposes an IoT-enabled deep learning framework for automated crop disease detection and classification. The collected crop images are pre-processed using scaling, normalization, and data augmentation techniques to improve image quality and model generalization. For efficient disease detection, a lightweight attention-driven convolutional neural network-based object detection approach, termed YOLO10-YSTs (You Only Look Once 10-Yield Surveillance Technology), is employed. The framework integrates the real-time detection capability of YOLO10 with an attention mechanism to emphasize disease-relevant regions and a stigmergy-based encoding strategy to enhance feature representation and extraction. In addition, a novel DBGU-LSTM (Dual-Bi-Gated Recurrent Unit with Long Short-Term Memory) architecture is developed for fine-grained crop disease classification. Results: The experimental results demonstrate that the proposed framework provides effective disease detection, discriminative feature representation, and fine-grained classification. The DBGU-LSTM architecture achieves improved classification performance and training stability compared with the conventional deep learning models considered in the study. Overall, the proposed approach achieves a classification accuracy of 97.99%, demonstrating its effectiveness for automated crop disease recognition. Discussion and conclusion: The results indicate that combining IoT-based environmental monitoring with attention-driven deep learning and recurrent feature-learning mechanisms can provide an effective solution for intelligent crop disease detection. The proposed framework enables accurate identification of disease patterns while supporting efficient and automated agricultural monitoring. Therefore, the approach has potential to assist farmers in early disease diagnosis, reduce crop losses, and improve agricultural productivity.
- Open Access
- 12 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cognitive and Developmental Systems
- Vol. PP  (23798920) , 1-16
A Developmentally-Staged Transformer Framework for Progressive Cognitive Learning in Language Understanding
Human language understanding develops progressively through stages of cognitive maturation, characterized by increasing memory capacity, attentional control, and abstraction. Inspired by this trajectory, we propose the Developmentally-Staged Transformer for Progressive Cognitive Learning (DST-PCL), a novel framework that models language understanding through staged architectural and functional growth. DST-PCL simulates cognitive development by incrementally expanding Transformer depth, input memory span, and attention complexity across a curriculum-aligned training schedule. Each stage activates a subset of model layers and imposes entropy-regularized attention constraints to reflect bounded human focus. A readiness-based transition controller governs stage advancement based on validation accuracy and attention sharpness. Theoretical analyses guarantee hypothesis class expansion, bounded generalization error, and stability under stage transitions. Empirical evaluations across three benchmark datasets BoolQ, MultiNLI, and ReClor demonstrate consistent gains over strong baselines, with up to +5.1% accuracy and +6.3 F1 improvement. Further, attention entropy and interpretability metrics confirm the cognitively plausible progression of model behavior. DST-PCL offers a new paradigm for cognitively grounded NLP systems that unify performance, developmental structure, and interpretability.
- 12 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Cognitive and Developmental Systems
- Vol. PP  (23798920) , 1-16
Open AccessA Developmentally-Staged Transformer Framework for Progressive Cognitive Learning in Language Understanding
Human language understanding develops progressively through stages of cognitive maturation, characterized by increasing memory capacity, attentional control, and abstraction. Inspired by this trajectory, we propose the Developmentally-Staged Transformer for Progressive Cognitive Learning (DST-PCL), a novel framework that models language understanding through staged architectural and functional growth. DST-PCL simulates cognitive development by incrementally expanding Transformer depth, input memory span, and attention complexity across a curriculum-aligned training schedule. Each stage activates a subset of model layers and imposes entropy-regularized attention constraints to reflect bounded human focus. A readiness-based transition controller governs stage advancement based on validation accuracy and attention sharpness. Theoretical analyses guarantee hypothesis class expansion, bounded generalization error, and stability under stage transitions. Empirical evaluations across three benchmark datasets BoolQ, MultiNLI, and ReClor demonstrate consistent gains over strong baselines, with up to +5.1% accuracy and +6.3 F1 improvement. Further, attention entropy and interpretability metrics confirm the cognitively plausible progression of model behavior. DST-PCL offers a new paradigm for cognitively grounded NLP systems that unify performance, developmental structure, and interpretability.
- 12 August 2025
- Published by IGI Global Scientific Publishing
- p. 293-318
QINN-Based Approach to Detect Anomalies in High-Dimensional Secure Data
Anomaly discovery is a crucial aspect of modern data analysis to finding unusual trends or behaviours in datasets across various domains like cybersecurity and finance and the healthcare. However, overfitting, in which the model becomes overly adapted to training data, results in false negatives and misclassifications, makes it difficult to target optimal detection capabilities. To get around this, training data must be carefully sanitized, eliminating unknown and irregular anomalous instances to guarantee precise anomaly detection. It is already difficult to accomplish optimal and significant feature extraction when working with noisy, high-dimensional data, like that found in network traffic. When features are too similar, overfitting and incorrect classifications may occur. However, classification accuracy may suffer if important features are removed. By improving likelihood estimation and feature separability, unsupervised techniques can assist in finding and keeping pertinent features to enhance model performance.
- 12 August 2025
- Published by IGI Global Scientific Publishing
- p. 293-318
QINN-Based Approach to Detect Anomalies in High-Dimensional Secure Data
Anomaly discovery is a crucial aspect of modern data analysis to finding unusual trends or behaviours in datasets across various domains like cybersecurity and finance and the healthcare. However, overfitting, in which the model becomes overly adapted to training data, results in false negatives and misclassifications, makes it difficult to target optimal detection capabilities. To get around this, training data must be carefully sanitized, eliminating unknown and irregular anomalous instances to guarantee precise anomaly detection. It is already difficult to accomplish optimal and significant feature extraction when working with noisy, high-dimensional data, like that found in network traffic. When features are too similar, overfitting and incorrect classifications may occur. However, classification accuracy may suffer if important features are removed. By improving likelihood estimation and feature separability, unsupervised techniques can assist in finding and keeping pertinent features to enhance model performance.
- 23 April 2025
- Published by Springer Nature in Data Mining and Knowledge Discovery
- Vol. 39  (4) , 1-38
Clustering using GRASP and path relinking for the max k-cut problem
- 23 April 2025
- Published by Springer Nature in Data Mining and Knowledge Discovery
- Vol. 39  (4) , 1-38
Clustering using GRASP and path relinking for the max k-cut problem
- 20 December 2024
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 38  (5) , 3148-3167
Adaptive Compression and Reconstruction for Multidimensional Medical Image Data: A Hybrid Algorithm for Enhanced Image Quality
Spatial regions within images typically hold greater priority over adjacent areas, especially in the context of medical images (MI) where minute details can have significant clinical implications. This research addresses the challenge of compressing medical image dimensions without compromising critical information by proposing an adaptive compression algorithm. The algorithm integrates a modified image enhancement module, clustering-based segmentation, and a variety of lossless and lossy compression techniques. Edge enhancement contrast limited adaptive histogram equalization (EE-CLAHE) and 2D adaptive anisotropic diffusion filter are employed to enhance and denoise the images, followed by adaptive expectation maximization clustering (AEMC) for segmentation into regions of interest (ROI) and non-ROI. The clustering process is optimized utilizing fuzzy c-means (FCM) and Otsu thresholding. Subsequently, distinct compression schemes are applied to ROI and non-ROI regions, such as Coiflet + Haar, Coiflet + Daubecheis, modified SPIHT Huffman, EZW, and SPIHT algorithms, to ensure effective storage and transmission while preserving diagnostic details. Experimental results demonstrate that the combination of the modified SPIHT Huffman algorithm for ROI and EZW for non-ROI yields superior reconstruction quality across various measures, enabling comprehensive analysis of multi-dimensional images from MRI, CT, and X-ray modalities.
- 20 December 2024
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 38  (5) , 3148-3167
Adaptive Compression and Reconstruction for Multidimensional Medical Image Data: A Hybrid Algorithm for Enhanced Image Quality
Spatial regions within images typically hold greater priority over adjacent areas, especially in the context of medical images (MI) where minute details can have significant clinical implications. This research addresses the challenge of compressing medical image dimensions without compromising critical information by proposing an adaptive compression algorithm. The algorithm integrates a modified image enhancement module, clustering-based segmentation, and a variety of lossless and lossy compression techniques. Edge enhancement contrast limited adaptive histogram equalization (EE-CLAHE) and 2D adaptive anisotropic diffusion filter are employed to enhance and denoise the images, followed by adaptive expectation maximization clustering (AEMC) for segmentation into regions of interest (ROI) and non-ROI. The clustering process is optimized utilizing fuzzy c-means (FCM) and Otsu thresholding. Subsequently, distinct compression schemes are applied to ROI and non-ROI regions, such as Coiflet + Haar, Coiflet + Daubecheis, modified SPIHT Huffman, EZW, and SPIHT algorithms, to ensure effective storage and transmission while preserving diagnostic details. Experimental results demonstrate that the combination of the modified SPIHT Huffman algorithm for ROI and EZW for non-ROI yields superior reconstruction quality across various measures, enabling comprehensive analysis of multi-dimensional images from MRI, CT, and X-ray modalities.
- Open Access
- 25 November 2024
- Published by Springer Nature in Research Square
Deception-Based Data Protection Using Multifile Encryption and Honey Diffusion
The importance of encryption algorithms has greatly increased as an increasing amount of communication and storage are becoming digital. Typical encryption algorithms operate by encrypting a file or message with a key which is distributed to trusted parties. However, if a malicious or untrustworthy individual is able to discover the key, any data protected by these algorithms becomes compromised. To address this, deception-based algorithms have been suggested, where false or meaningless data is mixed together with real data so that malicious parties will be unable to determine whether any data they have managed to obtain is legitimate. In this paper, we present two novel deception-based algorithms, each designed to address the other's limitations. The first algorithm uses a multifile encryption algorithm that merges multiple files, each with their own key, into a single generator file that outputs one of the original files when given their corresponding password. The second algorithm, Honey Diffusion, uses a diffusion model that generates a near-exact replica of the real data when given the correct password, and similar but distinct data when given an incorrect password. We also demonstrate some use cases of these algorithms.
- 25 November 2024
- Published by Springer Nature in Research Square
Open AccessDeception-Based Data Protection Using Multifile Encryption and Honey Diffusion
The importance of encryption algorithms has greatly increased as an increasing amount of communication and storage are becoming digital. Typical encryption algorithms operate by encrypting a file or message with a key which is distributed to trusted parties. However, if a malicious or untrustworthy individual is able to discover the key, any data protected by these algorithms becomes compromised. To address this, deception-based algorithms have been suggested, where false or meaningless data is mixed together with real data so that malicious parties will be unable to determine whether any data they have managed to obtain is legitimate. In this paper, we present two novel deception-based algorithms, each designed to address the other's limitations. The first algorithm uses a multifile encryption algorithm that merges multiple files, each with their own key, into a single generator file that outputs one of the original files when given their corresponding password. The second algorithm, Honey Diffusion, uses a diffusion model that generates a near-exact replica of the real data when given the correct password, and similar but distinct data when given an incorrect password. We also demonstrate some use cases of these algorithms.
- 9 January 2024
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 170, 107977
A novel attention-based cross-modal transfer learning framework for predicting cardiovascular disease
- 9 January 2024
- Published by Elsevier in Computers in Biology and Medicine
- Vol. 170, 107977
A novel attention-based cross-modal transfer learning framework for predicting cardiovascular disease
- 1 December 2023
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 83  (18) , 54249-54278
A context-sensitive multi-tier deep learning framework for multimodal sentiment analysis
- 1 December 2023
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 83  (18) , 54249-54278
A context-sensitive multi-tier deep learning framework for multimodal sentiment analysis
- 15 May 2023
- Published by Springer Nature in Soft Computing
- Vol. 27  (19) , 14189-14203
An integrated network topology and deep learning model for prediction of Alzheimer disease candidate genes
- 15 May 2023
- Published by Springer Nature in Soft Computing
- Vol. 27  (19) , 14189-14203
An integrated network topology and deep learning model for prediction of Alzheimer disease candidate genes