Author ORCID Identifier
Biosketch
K. Ramasamy receivedhis M.Sc. and M.Phil. degrees from Bharathidasan University, and his M.Tech degree from SASTRA Deemed University.He qualified the UGC - National Eligibility Test (NET) in the year 2013. He completed his Ph.D in Image Processing Techniquesin July 2026. He is currently working as an Assistant Professor in the Department of Electronics & Communication Engineering, Srinivasa Ramanujan Centre, SASTRA Deemed University, Kumbakonamsince 2006. He is a member of ISTE and ISSE. His research focuses on palmprint image analysis and machine learning, with an emphasis on developing robust deep learning pipelines for palmprint classification tasks. His areas of interest are Image Processing, Machine Learning and Deep Learning. He has published eleven articles in journals indexed in Scopus and SCI-E.
Date of Award
24-7-2026
Document Type
Thesis
School
School of Electrical & Electroncis Engineering
Programme
Ph.D.-Doctoral of Philosophy
First Advisor
Dr.A.Srinivasan
Keywords
Palmprint, Vision Transformer, Wavelet Transform, Regression, Diabetes
Abstract
Type 2 diabetes is one among the diabetes mellitus which has strong relation with lifestyle, genetics and metabolic dysfunction. The increasing global prevalence, rising healthcare costs, and high percentage of undiagnosed individuals leads to the need of noninvasive, cost-effective early detection methods. In conventional diagnostic methods, laboratory investigations and periodic clinical visits are needed, which may not be feasible for working and middle-aged populations. To overcome this, the research work proposes an automatic, non-invasive and early prediction of Type 2 diabetes using palmprint image analysis integrated with clinical data for identifying diabetic and non-diabetic individuals. The first proposed method aims to develop a palmprint image-based classification using Semi-Global Matching (SGM) combined with K-Means clustering for segmentation, Scale Invariant Feature Transform (SIFT) for feature extraction, and a Multiscale Deep Convolutional Neural Network (MSDCNN) for classification.
The performance of proposed MSDCNN method is analysed using PolyU dataset and real datasets by comparing with conventional methods. The proposed MSDCNN method achieved an accuracy of 94.33% on the PolyU dataset and 94.67% accuracy on a real-time dataset of 750 palmprint images. The objective of the second proposed method is to enhance the palmprint images in capturing fine-grained variations across palm layers and to improve the accuracy of prediction in identifying Type 2 diabetes. The proposed method employs Transverse Dyadic Wavelet Transform (TDyWT) and optimized Vision Transformer (ViT) are employed to capture texture, edge, and orientation-based features effectively while preserving spatial alignment.
Statistical and texture features such as mean, median, entropy, standard deviation, contrast, correlation, homogeneity, and energy are extracted from the optimized ViT for further processing. Also, Gabor filters are utilized to derive ridge density, average ridge density, and total angle deviation, capturing structural variations in ridges of palm images. Thus, the second proposed method enhances the palm image and extracts eleven discriminative features for predictive modeling. The third proposed method integrates the eleven features extracted from optimized ViT and Gabor filter with clinical data and regression models namely Bayesian-Optimized Multilevel Regression (BO-MLR) and Cubic Support Vector Machine (CSVM) aids for final prediction. Bayesian optimization enhances model generalization by tuning regression hyperparameters within a probabilistic framework, while Cubic SVM employs a third-degree polynomial kernel to model nonlinear decision boundaries.
Four methods are proposed namely Particle swarm optimized ViT with Bayesian optimized Multilevel regression (BOPS), Adam ViT with Bayesian optimized Multilevel regression (BOAD), Particle swarm optimized ViT with Cubic SVM (CKPS) and Adam optimized ViT with Cubic SVM (CKAD) and are evaluated using real palmprint images captured through a palmprint scanner system. Among the proposed methods, the Cubic-SVM with Adam-optimized Vision Transformer (CKAD) achieved superior results with 98.80% accuracy, 98.45% precision, 98.75% sensitivity, and 98.83% specificity. In overall, among the proposed methods CKAD achieved highest accuracy followed by CKPS, BOAD, BOPS and MSDCNN. The integration of optimized deep learning feature extraction with regression models achieved superior predictive performance compared to traditional methods.
Recommended Citation
K, Ramasamy Mr, "Application of Palmprint Data for Prediction of Diabetes" (2026). Theses and Dissertations. 228.
https://knowledgeconnect.sastra.edu/theses/228