The Dense Convolutional Network for Rice Disease Recognition Based on Transfer Learning

Abstract

Crop diseases significantly impact agricultural productivity, making their timely detection and control critical. This study leverages artificial intelligence to detect and identify plant diseases from images, aiming to improve the targeting of disease management and reduce chemical usage, thus mitigating contamination and enhancing agricultural sustainability.

Focusing on rice as a case study, we propose a novel method to detect three common rice diseases using a Transfer Learning-based Dense Convolutional Network model. To address the challenges in intelligent disease detection, we enhance the dataset through channel transformation, symmetry, and rotation, which augment the model's ability to generalize across diverse rice disease conditions. By utilizing deep convolutional neural network, the model automatically learns discriminative features, bypassing the need for traditional, labor-intensive manual feature extraction.

The Transfer Learning approach allows us to transfer knowledge learned from the ImageNet dataset to the rice disease detection task, significantly improving performance and reducing training time. We apply digital image processing techniques and CNN to construct the system, and use a Softmax classifier to handle the multiclass disease classification problem.

Extensive experiments comparing different models demonstrate that the DenseNet based Transfer Learning model outperforms traditional methods in rice disease detection. Specifically, the model achieves 100% accuracy on both training and testing datasets for detecting diseased rice leaves. Furthermore, 10-fold cross-validation results in a mean accuracy of 99.81%, confirming the robustness and validity of the model. Our approach offers high diagnostic accuracy and shows potential for scaling to diagnose various plant diseases, with significant implications for advancing agricultural practices.