A Lightweight DenseNet-121 with Efficient Channel Attention and Progressive Resizing for Rice Leaf Disease Detection

Read full text Summarise with VIRAT
Authors: Parthasarathi Jayaraman, Muthulakshmi Palaniyandi
Cited by: 0
References: 22

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Rice leaf diseases are a major threat to agricultural productivity.Early disease identification significantly improves yield stability.Automated diagnosis with good accuracy and efficiency is in high demand.Several Deep Learning architectures have been utilized for rice leaf identification, among them, DenseNet variants provide strong feature reuse capabilities but often struggle to capture subtle morphological variations of lesions.To improve the DenseNet model further, this work proposes a hybrid framework, where a lightweight DenseNet-121 model is integrated with the Efficient Channel Attention (ECA) module and a two-phase progressive resize strategy.The ECA module allows the model to selectively emphasize disease relevant features without adding much computational cost.The progressive resizing mechanism enhances spatial discrimination by gradually increasing the input resolution from 224→320→384 pixels.This strategy improves understanding subtle lesion boundaries that have been overlooked by standard single-scale training.The proposed DenseNet-121+ECA model is applied on a publicly available RiceLeafV3 dataset.It achieved an accuracy of 98.17%, representing an absolute improvement of 6% over the DenseNet-121 backbone model with minimum use of 7.05 million parameters, along with a 28 MB checkpoint footprint, facilitating deployment in edge devices.Results were compared with the conventional Transfer Learning models and existing literature, such that the model outperforms both in terms of accuracy and efficiency.Comprehensive evaluation using statistical significance testing, confusion matrix analysis, and Grad-CAM visualizations confirms the robustness, interpretability, and efficiency of the proposed approach for precision agriculture applications.

Fields of Study & Subjects

Smart Agriculture and AI Advanced Neural Network Applications Industrial Vision Systems and Defect Detection

Publication details

Journal / source
International journal of intelligent engineering and systems
Volume
19
Issue
4
Pages
126–147
Publication date
23 March 2026
Language
EN