Akshatha Prabhu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Akshatha Prabhu has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
- N. Shobha Rani
Department of Artificial Intelligence and Data Science, MURTI Research Centre, Smart Agriculture Laboratory, GITAM School of Technology, GITAM (Deemed to be University), Bengaluru, India
3 shared publications - Manikandaprabhu Perumalsamy
Amrita School of Computing, Amrita Vishwa Vidyapeetham
2 shared publications Achala Sharma J
1 shared publication- Anitha Raghavendra
Akshaya Designs and Solutions
1 shared publication - Anuvarshini Palaniswamy
Department of Computer Science, School of Computing, Amrita Vishwa Vidyapeetham, Mysuru Campus, India
1 shared publication Ashwini M V
1 shared publication
Latest Publications
- 6 October 2026
- Published by Springer Nature
- p. 158-173
An Automated Computer Vision Framework for Mango Weight Estimation Using Interactive Regression Models in Fruit Quality Management
- 6 October 2026
- Published by Springer Nature
- p. 158-173
An Automated Computer Vision Framework for Mango Weight Estimation Using Interactive Regression Models in Fruit Quality Management
- Open Access
- 23 September 2026
- Published by Springer Nature in Discover Artificial Intelligence
- Vol. 6 (1) , 1224
An explainable machine learning framework for early student dropout risk prediction and stratification
- 23 September 2026
- Published by Springer Nature in Discover Artificial Intelligence
- Vol. 6 (1) , 1224
Open AccessAn explainable machine learning framework for early student dropout risk prediction and stratification
- Open Access
- 23 March 2026
- Published by Elsevier in SSRN Electronic Journal
CLARE-NET: A Validation-Driven Approach Adaptive Ensemble Framework for Reliable Multi-class Retinal Disease Classification
Introducing a multi-class retinal disease classification adaptive ensemble method that is built on validation to classify colour fundus images. The framework combines complementary EfficientNet backbones with probability-level adaptive fusion and class-specific threshold optimization to enhance balanced multi-class performance in the presence of unbalanced clinical data distributions. The calibration of the representation of the retinal features is performed with the help of circular region-of-interest masking and contrast-limited adaptive histogram equalization (CLAHE). The principle of ensemble weights lies in the validation macro-F1 performance, where a single architecture is dynamically added without the complexity of the architecture.To standardize the representation of the retinal features, circular region-of-interest masking and contrast-limited adaptive histogram equalization (CLAHE) are used. Ensemble weights are based on validation macro-F1 performance, allowing the dynamic addition of individual models without introducing architectural complexity. The CLARE-NET framework is focused on computational efficiency, probability, reliability and the ability to deploy population-scale screenings. Adaptive probability-level-weighting of the ensemble based on macro-F1 performance of validation.Class-specific threshold optimization Threshold optimization involving minorities.Dual-backbone lightweight and non-stacking feature and graph-based fusion.
- 23 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCLARE-NET: A Validation-Driven Approach Adaptive Ensemble Framework for Reliable Multi-class Retinal Disease Classification
Introducing a multi-class retinal disease classification adaptive ensemble method that is built on validation to classify colour fundus images. The framework combines complementary EfficientNet backbones with probability-level adaptive fusion and class-specific threshold optimization to enhance balanced multi-class performance in the presence of unbalanced clinical data distributions. The calibration of the representation of the retinal features is performed with the help of circular region-of-interest masking and contrast-limited adaptive histogram equalization (CLAHE). The principle of ensemble weights lies in the validation macro-F1 performance, where a single architecture is dynamically added without the complexity of the architecture.To standardize the representation of the retinal features, circular region-of-interest masking and contrast-limited adaptive histogram equalization (CLAHE) are used. Ensemble weights are based on validation macro-F1 performance, allowing the dynamic addition of individual models without introducing architectural complexity. The CLARE-NET framework is focused on computational efficiency, probability, reliability and the ability to deploy population-scale screenings. Adaptive probability-level-weighting of the ensemble based on macro-F1 performance of validation.Class-specific threshold optimization Threshold optimization involving minorities.Dual-backbone lightweight and non-stacking feature and graph-based fusion.
- 2 January 2026
- Published by Springer Nature
- p. 123-147
Mango Ripeness Evaluation: A Hierarchical Deep Learning Approach
- 2 January 2026
- Published by Springer Nature
- p. 123-147
Mango Ripeness Evaluation: A Hierarchical Deep Learning Approach
- Open Access
- 9 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 213578-213599
FeaFusion-PomoNet: A Feature Fusion Driven Regression Framework for Non-Destruction Weight Estimation of Pomegranates
Non-destructive fruit weight estimation plays a critical role in precision agriculture, particularly for yield forecasting, grading, and logistics optimization. FeaFusion-PomoNet, a feature fusion-based regression framework, is proposed for estimating the weight of fruits using a self-collected image dataset comprising samples captured from multiple orientations. Handcrafted features—including texture, shape, geometric attributes, and pixel density—are extracted and optimized via the Boruta feature selection algorithm. Multiple regression models are evaluated, with Multiple Linear Regression (MLR) achieving the best performance, yielding R² scores of 0.97 and 0.92 on 80-20 and 70-30 train-test splits, respectively. Validation through Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) produces average relative error rates of 7.6% and 8.3%, indicating high predictive accuracy. A Pearson correlation coefficient of 0.99 confirms the robustness of the model. The method also supports orientation-independent image acquisition, making it adaptable to irregularly shaped fruits. Future work will focus on scaling the dataset, incorporating field-captured images, and improving robustness under real-world conditions, contributing to automated, cost-effective post-harvest systems.
- 9 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 213578-213599
Open AccessFeaFusion-PomoNet: A Feature Fusion Driven Regression Framework for Non-Destruction Weight Estimation of Pomegranates
Non-destructive fruit weight estimation plays a critical role in precision agriculture, particularly for yield forecasting, grading, and logistics optimization. FeaFusion-PomoNet, a feature fusion-based regression framework, is proposed for estimating the weight of fruits using a self-collected image dataset comprising samples captured from multiple orientations. Handcrafted features—including texture, shape, geometric attributes, and pixel density—are extracted and optimized via the Boruta feature selection algorithm. Multiple regression models are evaluated, with Multiple Linear Regression (MLR) achieving the best performance, yielding R² scores of 0.97 and 0.92 on 80-20 and 70-30 train-test splits, respectively. Validation through Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) produces average relative error rates of 7.6% and 8.3%, indicating high predictive accuracy. A Pearson correlation coefficient of 0.99 confirms the robustness of the model. The method also supports orientation-independent image acquisition, making it adaptable to irregularly shaped fruits. Future work will focus on scaling the dataset, incorporating field-captured images, and improving robustness under real-world conditions, contributing to automated, cost-effective post-harvest systems.
- 14 February 2024
- Published by Springer Nature
- p. 49-63
Handwritten Character Recognition for South Indian Languages Using Deep Learning
- 14 February 2024
- Published by Springer Nature
- p. 49-63
Handwritten Character Recognition for South Indian Languages Using Deep Learning
- Open Access
- 7 November 2023
- Published by Elsevier in Data in Brief
- Vol. 51, 109778
AMDPWE: Alphonso Mango Dataset for Precision Weight Estimation
- 7 November 2023
- Published by Elsevier in Data in Brief
- Vol. 51, 109778
Open AccessAMDPWE: Alphonso Mango Dataset for Precision Weight Estimation