Sharmila Gopalswamy
Research Subjects
Sharmila Gopalswamy has a total of 6 co-authors. Here are the top co-authors by number of shared publications:
- Kavitha Rajamohan
Department of Computer Science, Christ University, Bangalore
3 shared publications - Jayasankar Uthayakumar
Genesys Academy of Computer Science, Puducherry, India
1 shared publication - Kavitha Rajamohan
Department of Computer Science CHRIST (Deemed to University
1 shared publication - Lubin Balasubramanian1 shared publication
- Niranjana Kumara M1 shared publication
- Srinivasan Ramasamy
Senior Scientist (Soil Science) National Bureau of Soil Survey and Land Use Planning
1 shared publication
Latest Publications
- 14 October 2026
- Published by VFAST Research Platform in VFAST Transactions on Software Engineering
- Vol. 14 (4) , 156-169
Comprehensive Learning-Based Automatic Speaker Recognition Using Log-Mel Spectral Representation and Deep Neural Networks
Speaker recognition is a method for automatically recognizing a speaker from a recording of their speech or voice utterance. Speaker recognition technology has developed into a low-cost and dependable technique for person identification and verification. Deep learning (DL) methods are demonstrating promising outcomes for various speech-related applications. With the development of end-to-end models, DL has transformed the domain of Automatic Speech Recognition (ASR). These models can automatically learn high-level and discriminative representations directly from speech data without broad manual feature engineering. This study introduces a Comprehensive Learning-Based Automatic Speaker Recognition Framework using Deep Learning Architecture (CLASR-DLA). The primary objective of the proposed CLASR-DLA approach is to accurately classify and identify speakers from speech signals. Initially, the model performs preprocessing to enhance the quality and consistency of the input speech data. Subsequently, Log-Mel spectrogram representations are generated to extract more informative and discriminative acoustic features. In addition, a stacked long short-term memory network (SLSTM) is employed for speaker detection and classification. To further enhance model stability and generalization performance, the model parameters are optimized using the AdaBelief optimizer. Extensive simulations are conducted using benchmark dataset. The results demonstrate the improvement of the CLASR-DLA model under various evaluation metrics.
- 14 October 2026
- Published by VFAST Research Platform in VFAST Transactions on Software Engineering
- Vol. 14 (4) , 156-169
Comprehensive Learning-Based Automatic Speaker Recognition Using Log-Mel Spectral Representation and Deep Neural Networks
Speaker recognition is a method for automatically recognizing a speaker from a recording of their speech or voice utterance. Speaker recognition technology has developed into a low-cost and dependable technique for person identification and verification. Deep learning (DL) methods are demonstrating promising outcomes for various speech-related applications. With the development of end-to-end models, DL has transformed the domain of Automatic Speech Recognition (ASR). These models can automatically learn high-level and discriminative representations directly from speech data without broad manual feature engineering. This study introduces a Comprehensive Learning-Based Automatic Speaker Recognition Framework using Deep Learning Architecture (CLASR-DLA). The primary objective of the proposed CLASR-DLA approach is to accurately classify and identify speakers from speech signals. Initially, the model performs preprocessing to enhance the quality and consistency of the input speech data. Subsequently, Log-Mel spectrogram representations are generated to extract more informative and discriminative acoustic features. In addition, a stacked long short-term memory network (SLSTM) is employed for speaker detection and classification. To further enhance model stability and generalization performance, the model parameters are optimized using the AdaBelief optimizer. Extensive simulations are conducted using benchmark dataset. The results demonstrate the improvement of the CLASR-DLA model under various evaluation metrics.
- 28 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1555-1562
AgroSSL-Net: A Self-Supervised Deep Learning Framework for Crop Health Monitoring
Accurate crop health monitoring is essential for early disease identification, pest infestation detection, and nutrient deficiency assessment. Traditional deep learning (DL) models struggle with feature generalization due to limited labeled agricultural datasets. This study proposes a model AgroSSL-Net, a new self-supervised learning (SSL) framework to enhance feature learning from unlabeled crop images. This novel Modality-aware Contrastive Learning Network MoCLRNet method leverages contrastive learning and momentum-based representation learning to extract robust features. A fine-tuned CNN backbone with multi-head self-attention (MHSA) is employed for feature refinement, ensuring improved classification accuracy. Experiments on benchmark agricultural datasets demonstrate that MoCLRNet outperforms existing models, achieving an accuracy of 92.1% for disease classification, 91.8% for pest detection, and 90.6% for nutrient deficiency assessment. The proposed method also improves feature extraction efficiency, reducing computational complexity while enhancing robustness. These findings imply that AgroSSL-Net is an efficient approach for real-time, automated crop health monitoring in smart agriculture.
- 28 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1555-1562
AgroSSL-Net: A Self-Supervised Deep Learning Framework for Crop Health Monitoring
Accurate crop health monitoring is essential for early disease identification, pest infestation detection, and nutrient deficiency assessment. Traditional deep learning (DL) models struggle with feature generalization due to limited labeled agricultural datasets. This study proposes a model AgroSSL-Net, a new self-supervised learning (SSL) framework to enhance feature learning from unlabeled crop images. This novel Modality-aware Contrastive Learning Network MoCLRNet method leverages contrastive learning and momentum-based representation learning to extract robust features. A fine-tuned CNN backbone with multi-head self-attention (MHSA) is employed for feature refinement, ensuring improved classification accuracy. Experiments on benchmark agricultural datasets demonstrate that MoCLRNet outperforms existing models, achieving an accuracy of 92.1% for disease classification, 91.8% for pest detection, and 90.6% for nutrient deficiency assessment. The proposed method also improves feature extraction efficiency, reducing computational complexity while enhancing robustness. These findings imply that AgroSSL-Net is an efficient approach for real-time, automated crop health monitoring in smart agriculture.
- Open Access
- 2 June 2025
- Published by Academic Science Publications and Distributions in International Journal of Environmental Sciences
- Vol. 11 (7s) , 161-180
Evaluating Soil Attributes and Recommendation of Crop Using KFSMOTEC Approach in CNN Model
In modern agriculture, precise soil analysis and crop prediction are pivotal for optimizing agricultural practices and ensuring sustainable food production. This paper focuses on a comprehensive approach that integrates soil pH and nutrient detection with subsequent crop prediction, harnessing advanced image processing and machine learning techniques. First the input soil image is preprocessed. After preprocessing, Active contour models are utilized to accurately delineate the soil regions within input images. This segmentation is a fundamental step in subsequent PH and soil nutrient detection. For soil nutrient detection, an Extreme Learning Machine has been employed with Ensemble Kernels (EK-ELM). The goal of developing this novel EK-ELM method is to classify nutrients by the result obtained by aggregating different kernel functions. All the kernel functions were optimized using a weighted Ensemble by prior knowledge. This robust machine learning algorithm effectively processes the segmented soil images, extracting essential nutrient information from the soil's visual properties. Furthermore, the novel approaches K-Fold cross validation & Synthetic Minority Over-Sampling Technique implemented with Convolutional Neural Network (KFSMOTECNN) which comes into play for crop prediction. KFSMOTECNN have demonstrated their prowess in image analysis tasks. Here, they analyze the detected soil pH and nutrient information alongside additional environmental factors. By learning intricate relationships between these variables, KFSMOTECNN facilitate accurate crop yield predictions, aiding farmers in selecting suitable crop varieties. The proposed KFSMOTECNN approach provides 99% accuracy, 98.5% sensitivity, 98.62% specificity, 98% precision, and 98.21% F1-score for real soil dataset. These results unequivocally establish the superiority of our proposed approach -based crop prediction method when compared to conventional approaches.
- 2 June 2025
- Published by Academic Science Publications and Distributions in International Journal of Environmental Sciences
- Vol. 11 (7s) , 161-180
Open AccessEvaluating Soil Attributes and Recommendation of Crop Using KFSMOTEC Approach in CNN Model
In modern agriculture, precise soil analysis and crop prediction are pivotal for optimizing agricultural practices and ensuring sustainable food production. This paper focuses on a comprehensive approach that integrates soil pH and nutrient detection with subsequent crop prediction, harnessing advanced image processing and machine learning techniques. First the input soil image is preprocessed. After preprocessing, Active contour models are utilized to accurately delineate the soil regions within input images. This segmentation is a fundamental step in subsequent PH and soil nutrient detection. For soil nutrient detection, an Extreme Learning Machine has been employed with Ensemble Kernels (EK-ELM). The goal of developing this novel EK-ELM method is to classify nutrients by the result obtained by aggregating different kernel functions. All the kernel functions were optimized using a weighted Ensemble by prior knowledge. This robust machine learning algorithm effectively processes the segmented soil images, extracting essential nutrient information from the soil's visual properties. Furthermore, the novel approaches K-Fold cross validation & Synthetic Minority Over-Sampling Technique implemented with Convolutional Neural Network (KFSMOTECNN) which comes into play for crop prediction. KFSMOTECNN have demonstrated their prowess in image analysis tasks. Here, they analyze the detected soil pH and nutrient information alongside additional environmental factors. By learning intricate relationships between these variables, KFSMOTECNN facilitate accurate crop yield predictions, aiding farmers in selecting suitable crop varieties. The proposed KFSMOTECNN approach provides 99% accuracy, 98.5% sensitivity, 98.62% specificity, 98% precision, and 98.21% F1-score for real soil dataset. These results unequivocally establish the superiority of our proposed approach -based crop prediction method when compared to conventional approaches.
- 4 July 2024
- Published by Springer Nature
- p. 133-144
Digital Soil Texture Classification Using Machine Learning Approaches
- 4 July 2024
- Published by Springer Nature
- p. 133-144
Digital Soil Texture Classification Using Machine Learning Approaches
- 15 August 2023
- Published by Springer Nature in International Journal of Information Technology
- Vol. 15 (7) , 3791-3805
Soil classification using active contour model for efficient texture feature extraction
- 15 August 2023
- Published by Springer Nature in International Journal of Information Technology
- Vol. 15 (7) , 3791-3805
Soil classification using active contour model for efficient texture feature extraction
- 1 July 2022
- Published by Springer Nature
- p. 333-354
A Systematic Literature Review on Image Preprocessing and Feature Extraction Techniques in Precision Agriculture
- 1 July 2022
- Published by Springer Nature
- p. 333-354
A Systematic Literature Review on Image Preprocessing and Feature Extraction Techniques in Precision Agriculture