International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Plant Disease Detection Using Dynamic Knowledge Distillation and Attention Mechanism

Authors :

Mohammad Ghasemi Arian¹ Mohammad Hossein Yaghmaee Moghaddam²

1- Ferdowsi university of mashhad 2- Ferdowsi university of mashhad

Keywords :

Knowledge Distillation،Teacher،Student،PlantDoc،Attention،ResNet

Abstract :

One of the key aspects of modern agriculture is the detection of plant diseases. With the significant advancement of technology, one of the simplest and most cost-effective methods for assessing plant health is through image-based data. A common approach for analyzing such data involves the use of deep convolutional neural networks (CNNs). However, due to the computationally intensive nature of these networks, deploying them on edge devices is often challenging and costly. In this work, we propose a dynamic knowledge distillation framework that transfers knowledge from a powerful teacher model based on ResNet-50 to a compact student model built upon ResNet-18. This framework adaptively regulates the strength of knowledge distillation based on simulated noise levels during training, improving robustness under diverse conditions without requiring explicit noise estimation at inference. Experiments conducted on the PlantDoc dataset demonstrate that our method improves the student model’s accuracy from 61.86% to 65.25%, which is 2.11% higher than the teacher model’s accuracy, while reducing the student model's size by 47.2 MB compared to the teacher. The proposed approach provides a practical solution for real-time and on-device plant disease diagnosis.