Muralikrishna Shamaiah Narayanarao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Muralikrishna Shamaiah Narayanarao has a total of 45 co-authors. Here are the top co-authors by number of shared publications:
- B. Ashwath Rao
Manipal Institute of technology, Manipal Academy of Higher Education
3 shared publications - Raghavendra Ganiga
Manipal Academy of Higher Education, Manipal Institute of Technology, Manipal, Karnataka, India
3 shared publications - Arun P. Parameswaran
Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal, Karnataka, India
2 shared publications - Aruna Doreen Manezes2 shared publications
- K. Raghurama Holla
Department of Computer Science, Mangalore University, Mangalore, Karnataka
2 shared publications - N. Harivinod
St Joseph Engineering College, Mangaluru, India
2 shared publications
Latest Publications
- Open Access
- 3 August 2026
- Published by Taylor & Francis in Materials Technology
- Vol. 41 (1)
High-performance one-dimensional defective photonic crystal biosensor featuring semiconductor defect layer and machine learning classification for early cancer diagnosis
Cancer ranks as the second most common cause of death worldwide, comprising a variety of diseases marked by irregular and unchecked cell growth. This pathological condition can impact almost any pa...
- 3 August 2026
- Published by Taylor & Francis in Materials Technology
- Vol. 41 (1)
Open AccessHigh-performance one-dimensional defective photonic crystal biosensor featuring semiconductor defect layer and machine learning classification for early cancer diagnosis
Cancer ranks as the second most common cause of death worldwide, comprising a variety of diseases marked by irregular and unchecked cell growth. This pathological condition can impact almost any pa...
- 2 July 2026
- Published by Springer Nature
- p. 391-401
Predictive Modeling of Stock Prices with ESG Events: An AI Approach for Indian Equity Markets
- 2 July 2026
- Published by Springer Nature
- p. 391-401
Predictive Modeling of Stock Prices with ESG Events: An AI Approach for Indian Equity Markets
- Open Access
- 15 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 90995-91022
Evaluation of Feature Selection Techniques for LSTM- and GRU-Based Lower Limb Joint Angle Prediction Using Multimodal Biomechanical Data
Accurate lower-limb joint-angle prediction is vital for gait rehabilitation and wearable assistive systems; however, the high dimensionality of multimodal biomechanical data poses a significant challenge for resource-limited settings. Efficient feature selection is therefore essential to reduce model complexity while preserving predictive performance. To date, no prior study has systematically compared feature selection methods for LSTM/GRU-based lower-limb joint-angle prediction using multimodal gait data. We evaluate Recursive Feature Elimination (RFE), LassoCV, Mutual Information (MI), Autoencoder, and Principal Component Analysis (PCA) based feature subsets to train LSTM/GRU models that predict hip, knee, and ankle angles across sagittal, frontal, and transverse planes under both random 80/20 split and Leave-One-Subject-Out (LOSO) validation settings. Experiments on 13 healthy adults (7 trials each; 91 trials total), normalized and resampled to 1001 samples per trial, show that under the random 80/20 split, Autoencoder-selected features improved random-split prediction performance, reducing average RMSE by up to 57.77% for raw features and 54.99% for combined raw-derived features relative to PCA. Friedman tests confirmed significant differences among methods for both feature settings (raw: χ2 = 19.111, p < 0.001; combined: χ2 = 23.911, p < 0.0001). However, LOSO validation provides a stricter assessment of subject-independent generalization because each subject is held out entirely for testing and excluded from the training fold. Under LOSO validation using combined raw-derived features, the performance trend changed, with LassoCV reducing the mean RMSE relative to Autoencoder by up to 31.26% for GRU and 32.05% for LSTM. By identifying the most informative variables, the proposed pipeline provides guidance for prioritizing input features and sensing modalities. Such prioritization may help lower data-acquisition requirements and simplify model inputs in future wearable and rehabilitation applications. However, embedded deployment, inference latency, memory footprint, parameter count, power consumption, and reduced-sensor hardware performance were not experimentally validated in the present study. The present evaluation was conducted on healthy-subject gait data, and future validation in clinical populations is warranted to assess generalizability.
- 15 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 90995-91022
Open AccessEvaluation of Feature Selection Techniques for LSTM- and GRU-Based Lower Limb Joint Angle Prediction Using Multimodal Biomechanical Data
Accurate lower-limb joint-angle prediction is vital for gait rehabilitation and wearable assistive systems; however, the high dimensionality of multimodal biomechanical data poses a significant challenge for resource-limited settings. Efficient feature selection is therefore essential to reduce model complexity while preserving predictive performance. To date, no prior study has systematically compared feature selection methods for LSTM/GRU-based lower-limb joint-angle prediction using multimodal gait data. We evaluate Recursive Feature Elimination (RFE), LassoCV, Mutual Information (MI), Autoencoder, and Principal Component Analysis (PCA) based feature subsets to train LSTM/GRU models that predict hip, knee, and ankle angles across sagittal, frontal, and transverse planes under both random 80/20 split and Leave-One-Subject-Out (LOSO) validation settings. Experiments on 13 healthy adults (7 trials each; 91 trials total), normalized and resampled to 1001 samples per trial, show that under the random 80/20 split, Autoencoder-selected features improved random-split prediction performance, reducing average RMSE by up to 57.77% for raw features and 54.99% for combined raw-derived features relative to PCA. Friedman tests confirmed significant differences among methods for both feature settings (raw: χ2 = 19.111, p < 0.001; combined: χ2 = 23.911, p < 0.0001). However, LOSO validation provides a stricter assessment of subject-independent generalization because each subject is held out entirely for testing and excluded from the training fold. Under LOSO validation using combined raw-derived features, the performance trend changed, with LassoCV reducing the mean RMSE relative to Autoencoder by up to 31.26% for GRU and 32.05% for LSTM. By identifying the most informative variables, the proposed pipeline provides guidance for prioritizing input features and sensing modalities. Such prioritization may help lower data-acquisition requirements and simplify model inputs in future wearable and rehabilitation applications. However, embedded deployment, inference latency, memory footprint, parameter count, power consumption, and reduced-sensor hardware performance were not experimentally validated in the present study. The present evaluation was conducted on healthy-subject gait data, and future validation in clinical populations is warranted to assess generalizability.
- Open Access
- 28 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of the Computer Society
- Vol. 7, 889-900
A Unified Framework for Age-Invariant Face Recognition Using GANs and Vision Transformers
Age-invariant face recognition (AIFR) remains a persistent challenge in computer vision due to significant facial changes over time. While existing methods tackle this problem using either generative age simulation, contrastive embedding alignment, or feature disentanglement independently, they fall short in generalization and identity preservation when evaluated across datasets. This limitation highlights the need for a unified framework that can combine the strengths of these complementary strategies into a cohesive pipeline. In this work, we propose a unified deep learning framework that integrates CycleGAN-based age simulation, contrastive learning, and deep feature extraction through multiple backbone architectures including ResNet, EfficientNet, and Vision Transformers (ViT). The proposed approach uses a Siamese network with shared weights, where each branch includes a frozen pretrained backbone and an MLP projection head, trained with contrastive loss for identity similarity estimation. This framework is designed to jointly achieve image realism, representation robustness, and cross-dataset generalization. Experiments on UTKFace, MORPH, and FG-NET datasets demonstrate the effectiveness of our method in learning robust, age-invariant facial representations. We show that models in the non-unified framework achieve 98.92% and 99.14% accuracy on MORPH, but their performance drops to 53.49% and 52.89% on FG-NET under zero-shot generalization. In contrast, the proposed unified framework maintains 80.55% and 80.08% accuracy, reducing the generalization gap by over 27% and demonstrating robust cross-age performance. The results highlight the need for a unified framework to achieve robust cross-dataset generalization in age-invariant face recognition.
- 28 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of the Computer Society
- Vol. 7, 889-900
Open AccessA Unified Framework for Age-Invariant Face Recognition Using GANs and Vision Transformers
Age-invariant face recognition (AIFR) remains a persistent challenge in computer vision due to significant facial changes over time. While existing methods tackle this problem using either generative age simulation, contrastive embedding alignment, or feature disentanglement independently, they fall short in generalization and identity preservation when evaluated across datasets. This limitation highlights the need for a unified framework that can combine the strengths of these complementary strategies into a cohesive pipeline. In this work, we propose a unified deep learning framework that integrates CycleGAN-based age simulation, contrastive learning, and deep feature extraction through multiple backbone architectures including ResNet, EfficientNet, and Vision Transformers (ViT). The proposed approach uses a Siamese network with shared weights, where each branch includes a frozen pretrained backbone and an MLP projection head, trained with contrastive loss for identity similarity estimation. This framework is designed to jointly achieve image realism, representation robustness, and cross-dataset generalization. Experiments on UTKFace, MORPH, and FG-NET datasets demonstrate the effectiveness of our method in learning robust, age-invariant facial representations. We show that models in the non-unified framework achieve 98.92% and 99.14% accuracy on MORPH, but their performance drops to 53.49% and 52.89% on FG-NET under zero-shot generalization. In contrast, the proposed unified framework maintains 80.55% and 80.08% accuracy, reducing the generalization gap by over 27% and demonstrating robust cross-age performance. The results highlight the need for a unified framework to achieve robust cross-dataset generalization in age-invariant face recognition.
- 20 February 2026
- Published by Springer Nature in Japanese Journal of Statistics and Data Science
- Vol. 9 (1) , 169-199
Environmental hazards of water scarcity: the role of Varthur Lake’s water quality in groundwater depletion
- 20 February 2026
- Published by Springer Nature in Japanese Journal of Statistics and Data Science
- Vol. 9 (1) , 169-199
Environmental hazards of water scarcity: the role of Varthur Lake’s water quality in groundwater depletion
- Open Access
- 31 December 2025
- Published by Atlantis Press SARL
- p. 293-304
From Pixels to Words: ResNet–LSTM Based Image Captioning with Greedy and Beam Search
- 31 December 2025
- Published by Atlantis Press SARL
- p. 293-304
Open AccessFrom Pixels to Words: ResNet–LSTM Based Image Captioning with Greedy and Beam Search
- Open Access
- 31 December 2025
- Published by Atlantis Press SARL
- p. 322-331
Attention Guided Medical Image Captioning Using ResNet–LSTM
- 31 December 2025
- Published by Atlantis Press SARL
- p. 322-331
Open AccessAttention Guided Medical Image Captioning Using ResNet–LSTM
- Open Access
- 31 December 2025
- Published by Atlantis Press SARL
- p. 257-268
Poem-to-Music Retrieval through Multilingual Emotion Curves: A Low-Resource and Explainable Approach
- 31 December 2025
- Published by Atlantis Press SARL
- p. 257-268
Open AccessPoem-to-Music Retrieval through Multilingual Emotion Curves: A Low-Resource and Explainable Approach
- Open Access
- 21 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. PP, 1
Evaluation of LSTM, GRU and ANFIS Models for Ankle Angle and Ankle Moment Prediction Using Biomechanical Data
Accurate prediction of joint angles and moments is crucial for understanding human gait and developing assistive technologies, such as exoskeletons and prosthetics. This study compares the performance of Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRU), and Adaptive Neuro-Fuzzy Inference Systems (ANFIS) for predicting ankle angle and ankle moment at walking speeds of 1 km/h, 2 km/h, and 3 km/h. A multimodal sensor fusion approach was employed, incorporating electromyography (EMG), Ground Reaction Forces (GRF), and knee angle data as model inputs to enhance predictive accuracy. The performance of the models were assessed using Root Mean Square Error (RMSE) and Coefficient of Determination (R²) metrics, under both subject-specific and Leave-One-Subject-Out (LOSO) cross-subject validation frameworks. The results demonstrate that, under subject-specific evaluation, ANFIS achieved superior accuracy for ankle angle prediction at higher walking speeds, outperforming LSTM and GRU by up to 10.79% and 7.08% in RMSE at 3 km/h, respectively. Conversely, GRU provided better performance for ankle moment prediction under subject-specific evaluation, achieving up to 29.04% and 18.27% lower RMSE than ANFIS and LSTM at 1 km/h, respectively, highlighting its robustness under higher gait variability. Under the LOSO design, GRU consistently demonstrated superior generalization capability across subjects and walking speeds, reducing RMSE by 13.5% over ANFIS at 1 km/h and 2.84% at 3 km/h for ankle angle prediction, and achieving up to 17.5% and 4.5% lower RMSE compared to ANFIS and LSTM for ankle moment prediction. The comparative results were further validated using paired t-tests on RMSE values, confirming statistically significant performance differences between models. These findings highlight the importance of selecting appropriate modeling architectures for biomechanical prediction tasks and identify GRU as a promising candidate for robust gait estimation applicable to assistive and rehabilitative technologies.
- 21 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. PP, 1
Open AccessEvaluation of LSTM, GRU and ANFIS Models for Ankle Angle and Ankle Moment Prediction Using Biomechanical Data
Accurate prediction of joint angles and moments is crucial for understanding human gait and developing assistive technologies, such as exoskeletons and prosthetics. This study compares the performance of Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRU), and Adaptive Neuro-Fuzzy Inference Systems (ANFIS) for predicting ankle angle and ankle moment at walking speeds of 1 km/h, 2 km/h, and 3 km/h. A multimodal sensor fusion approach was employed, incorporating electromyography (EMG), Ground Reaction Forces (GRF), and knee angle data as model inputs to enhance predictive accuracy. The performance of the models were assessed using Root Mean Square Error (RMSE) and Coefficient of Determination (R²) metrics, under both subject-specific and Leave-One-Subject-Out (LOSO) cross-subject validation frameworks. The results demonstrate that, under subject-specific evaluation, ANFIS achieved superior accuracy for ankle angle prediction at higher walking speeds, outperforming LSTM and GRU by up to 10.79% and 7.08% in RMSE at 3 km/h, respectively. Conversely, GRU provided better performance for ankle moment prediction under subject-specific evaluation, achieving up to 29.04% and 18.27% lower RMSE than ANFIS and LSTM at 1 km/h, respectively, highlighting its robustness under higher gait variability. Under the LOSO design, GRU consistently demonstrated superior generalization capability across subjects and walking speeds, reducing RMSE by 13.5% over ANFIS at 1 km/h and 2.84% at 3 km/h for ankle angle prediction, and achieving up to 17.5% and 4.5% lower RMSE compared to ANFIS and LSTM for ankle moment prediction. The comparative results were further validated using paired t-tests on RMSE values, confirming statistically significant performance differences between models. These findings highlight the importance of selecting appropriate modeling architectures for biomechanical prediction tasks and identify GRU as a promising candidate for robust gait estimation applicable to assistive and rehabilitative technologies.
- Open Access
- 26 June 2025
- Published by American Chemical Society (ACS) in ACS Omega
- Vol. 10 (26) , 28362-28371
Q-Learning-Based Multivariate Nonlinear Model Predictive Controller: Experimental Validation on Batch Reactor for Temperature Trajectory Tracking
This study introduces a Q-learning-based nonlinear model predictive control (QL-NMPC) framework for temperature control in batch reactors. A reinforcement learning agent is trained in simulation to learn optimal control strategies using coolant flow rate and heater current as inputs. The resulting policy, represented as a Q-table, is implemented in real time on a physical reactor setup using the NVIDIA Jetson Orin platform. The proposed QL-NMPC framework employs a value iteration-based Q-learning algorithm, enabling model-free policy optimization without explicit policy evaluation steps, and demonstrates effective temperature tracking while highlighting the potential of reinforcement learning for controlling nonlinear batch processes without relying on system identification.
- 26 June 2025
- Published by American Chemical Society (ACS) in ACS Omega
- Vol. 10 (26) , 28362-28371
Open AccessQ-Learning-Based Multivariate Nonlinear Model Predictive Controller: Experimental Validation on Batch Reactor for Temperature Trajectory Tracking
This study introduces a Q-learning-based nonlinear model predictive control (QL-NMPC) framework for temperature control in batch reactors. A reinforcement learning agent is trained in simulation to learn optimal control strategies using coolant flow rate and heater current as inputs. The resulting policy, represented as a Q-table, is implemented in real time on a physical reactor setup using the NVIDIA Jetson Orin platform. The proposed QL-NMPC framework employs a value iteration-based Q-learning algorithm, enabling model-free policy optimization without explicit policy evaluation steps, and demonstrates effective temperature tracking while highlighting the potential of reinforcement learning for controlling nonlinear batch processes without relying on system identification.