GreenMedNet: A computationally inexpensive hyperparameter optimised dual attention based insulin leaves categorization approach - Université d'Évry
Communication Dans Un Congrès Année : 2024

GreenMedNet: A computationally inexpensive hyperparameter optimised dual attention based insulin leaves categorization approach

Résumé

Automated classification of distinct leaf categories through advanced deep neural technologies has emerged as a notable research trend. The primary aim of this research is to design 'GreenMedNet,' an advanced and computationally efficient deep neural network. This network is specifically tailored to automatically differentiate between diseased and healthy insulin leaves, based on their inherent hyperglycemic properties. This research is driven by the practical importance of the field. The primary contribution of our work lies in the creation of the novel 'GreenMedNet' architecture which is designed to effectively balance performance efficiency with a minimal total count of parameters. This balance is achieved through integrating Depth-wise Convolution, GlobalAveragePooling2D, Batch Normalization, and Dropout layers into the architecture of GreenMedNet. Depth-wise Convolution and GlobalAveragePooling2D layers play a crucial role in minimizing the total parameter count, whereas Batch Normalization and Dropout layers expedite convergence and mitigate overfitting, respectively. Additionally, the incorporation of a dual attention mechanism enhances the network's performance by selectively emphasizing important features both spatially and channel-wise. To support our research, we have curated a novel well-balanced database namely, 'Nature's Diabetes Helpers: Insulin Leaves Repository,' containing 429 images distributed between healthy and damaged classes exclusively which serves as another for this work. Notably diverse, this repository forms a unique resource. Our designed network demonstrates exceptional performance, achieving a Test Accuracy of 99.47%, along with high Precision, Recall, and F1 scores, which underscores its superiority over established networks.
Fichier non déposé

Dates et versions

hal-04853150 , version 1 (21-12-2024)

Identifiants

Citer

Ranadhir Das, Sriparna Banerjee, Bidisha Samanta, Swati Chowdhuri, Sheli Sinha Chaudhuri, et al.. GreenMedNet: A computationally inexpensive hyperparameter optimised dual attention based insulin leaves categorization approach. Fifth International Conference on Intelligent Data Science Technologies and Applications (IDSTA 2024), Sep 2024, Dubrovik, Croatia. pp.172-176, ⟨10.1109/IDSTA62194.2024.10747000⟩. ⟨hal-04853150⟩
0 Consultations
0 Téléchargements

Altmetric

Partager

More