Vadivu G

Prof. Vadivu G


Research Focus & Metrics

Academic Career Length

6Years (2021-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
24 July 2021 in Soft Computing
Verified
Journal Article
Closed access
PDF file
SSS. Sivasankari
S. Sivasankari
Vadivu GVadivu G
Vadivu G

13 March 2021 in Journal of Ambient Intelligence and Humanized Computing
Verified
Journal Article
Closed access
PDF file
ASA. Shanthini
A. Shanthini
GMGunasekaran Manogaran
Gunasekaran Manogaran
Vadivu GVadivu G
Vadivu G
KKKottilingam Kottursamy
Kottilingam Kottursamy
PNP. Nithyakani
P. Nithyakani
CFC. Fancy
C. Fancy

17 October 2024 in Veterinary Sciences
Verified
Journal Article
Open access
PDF file
CSChamirti Senthilkumar
Chamirti Senthilkumar
SSSindhu Chandra Sekharan
Sindhu Chandra Sekharan
Vadivu GVadivu G
Vadivu G
Suresh NeethirajanSuresh Neethirajan
Suresh Neethirajan
Lumpy Skin Disease (LSD) poses a significant threat to agricultural economies, particularly in livestock-dependent countries like India, due to its high transmission rate leading to severe morbidity and mortality among cattle. This underscores the urgent need for early and accurate detection to effectively manage and mitigate outbreaks. Leveraging advancements in computer vision and artificial intelligence, our research develops an automated system for LSD detection in cattle using deep learning techniques. We utilized two publicly available datasets comprising images of healthy cattle and those with LSD, including additional images of cattle affected by other diseases to enhance specificity and ensure the model detects LSD specifically rather than general illness signs. Our methodology involved preprocessing the images, applying data augmentation, and balancing the datasets to improve model generalizability. We evaluated over ten pretrained deep learning models—Xception, VGG16, VGG19, ResNet152V2, InceptionV3, MobileNetV2, DenseNet201, NASNetMobile, NASNetLarge, and EfficientNetV2S—using transfer learning. The models were rigorously trained and tested under diverse conditions, with performance assessed using metrics such as accuracy, sensitivity, specificity, precision, F1-score, and AUC-ROC. Notably, VGG16 and MobileNetV2 emerged as the most effective, achieving accuracies of 96.07% and 96.39%, sensitivities of 93.75% and 98.57%, and specificities of 97.14% and 94.59%, respectively. Our study critically highlights the strengths and limitations of each model, demonstrating that while high accuracy is achievable, sensitivity and specificity are crucial for clinical applicability. By meticulously detailing the performance characteristics and including images of cattle with other diseases, we ensured the robustness and reliability of the models. This comprehensive comparative analysis enriches our understanding of deep learning applications in veterinary diagnostics and makes a substantial contribution to the field of automated disease recognition in livestock farming. Our findings suggest that adopting such AI-driven diagnostic tools can enhance the early detection and control of LSD, ultimately benefiting animal health and the agricultural economy.


1 January 2023
Verified
Conference Paper
Open access
PDF file
SAShobanaDevi A
ShobanaDevi A
LPL. Anantha Padmanabhan
L. Anantha Padmanabhan
Vadivu GVadivu G
Vadivu G
Pao-Ann HsiungPao-Ann Hsiung
Pao-Ann Hsiung
English has now become a global language and is commonly spoken even amongst non-indigenous English societies. Grammar can be understood as an arrangement of words to form proper and meaningful sentences. English has a vast vocabulary and hence grammatical errors and typos are likely to happen for an average person. It is therefore important to have a medium that can help correct these errors and in doing so can help users improve their writing skills. The proposed work called Gramcorrector is an application that does both Grammar & spell correction. The availability of large volumes of data and good compute power paved the way for the development of such deep learning models. Gramcorrector is based on the Gramformer model, which uses the state of art transformer architecture. The core of this application is a Grammar & spell corrector model that has the capability to understand the intent of the sentence, identify all the possible errors and also apply accurate corrections. The corrected sentence or paragraph is displayed in the application UI along with the edits made on each sentence. This helps users to learn from their mistakes and also to improve his/her proficiency in English.

5 August 2025 in IEEE Access
Verified
Journal Article
Open access
JKJ Kavipriya
J Kavipriya
Vadivu GVadivu G
Vadivu G
Particularly in agricultural contexts where exact delineation of land cover types is vital for resource management and planning, semantic segmentation is a key technique for high-resolution image interpretation. Although U-Net-based designs have shown significant success in Land Use and Land Cover (LULC) identification, they sometimes struggle when segmenting classes with comparable spectral, textural, and intensity traits. Moreover, generating large-scale ground truth across wide agricultural areas is still time consuming and labor intensive, which greatly limits the creation of scalable segmentation solutions. This work presents a new architecture called Deep Pro Agri-UNet (DPA-UNet) meant specifically for agricultural field segmentation with high-resolution images sourced from Google Earth Pro in order to solve these problems. DPA-UNet improves the model’s ability to extract discriminative features and properly separate spectrally similar land classes by means of a multi-branch spatial and channel attention mechanism at the model’s bottleneck layer. A Dice loss function is included into the training goal to offset the consequences of class imbalance common in agricultural datasets. Attention gates are also included into the decoder path to selectively hone feature maps from the encoder to emphasize spatially relevant areas during the upsampling process. Experimental results show that DPA-UNet significantly outperforms traditional U-Net models, with an 81.8% total accuracy and an 82.8% Intersection over Union (IoU). While keeping a lower computational load, the proposed model efficiently lowers segmentation mistakes in heterogeneous agricultural areas. The findings confirm that DPA-UNet provides a scalable, precise, and computationally efficient solution for large scale agricultural monitoring, therefore supporting applications in sustainable land management, informed policy decision making, and precision farming.