Citation

Altabaji, Wassem Ibrahim Abdelhamid Eldusuqi (2025) Identification of Rice Leaf Diseases using Deep Learning for edge computing. Masters thesis, Multimedia University.

Full text not available from this repository.

Abstract

The rapid identification of rice leaf diseases is crucial. However, the deployment of low-cost edge hardware is constrained by the limitations in memory, power, and processing capabilities. This study addresses this gap by evaluating convolutional neural networks (CNNs) for diagnostic purposes on devices and applying systematic compression to meet real-time requirements. There were three main goals for this study, to use transfer learning to make a deep learning model that could detect and classify rice leaf diseases, to deal with the trade-off between speed and accuracy in disease identification, and to look into techniques to compress models for edge devices deployment. A unified protocol was employed to implement transfer learning on Xception and MobileNetV2. The RiceLeafNet, a compact, task-oriented CNN designed to serve as an effective baseline. The models were evaluated using both Google Colab, which offers extensive resources, and Raspberry Pi 5, which operated with limited computational resources. This comparison aimed to assess the accuracy, size, and prediction speed. Three compression methods were examined which are filter pruning, post-training quantisation (from 16-bit to 2-bit), and knowledge distillation, to clarify the trade-offs between accuracy, throughput, and footprint. Utilising 16-bit quantisation, MobileNetV2 maintained an accuracy of 67.88% while achieving 39.61 FPS at a size of 11.91 MB on the Raspberry Pi 5. The accuracy of RiceLeafNet remained at 73.37%, with a size of 9.25 MB and a performance of 17.12 FPS. The process of knowledge distillation resulted in smaller models as students, each with a uniform size of 4.83MB, while maintaining competitive accuracy and realtime throughput on the Raspberry Pi 5. The MobileNetV2 student proved 63.21% at 11.64 FPS, and the RiceLeafNet student achieved 67.07% at 11.34 FPS. The stated goals are achievable, including demonstrating that transfer learning yields productive classifiers, while quantisation and distillation address the equilibrium between speed and accuracy in optimised architectures. Additionally, to get the best performance, requires a carefully designed baseline model. RiceLeafNet is a strong practical solution and a reliable benchmark. Its accuracy and on-device efficiency are equal to or better than those of transfer-learned alternatives. This method makes it possible to use on edge devices with limited resources.

Item Type:	Thesis (Masters)
Additional Information:	Call No.: Q325.73 .A48 2025
Uncontrolled Keywords:	Deep learning (Machine learning)
Subjects:	Q Science > Q Science (General) > Q300-390 Cybernetics
Divisions:	Faculty of Artificial Intelligence & Engineering (FAIE)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	16 Apr 2026 02:16
Last Modified:	16 Apr 2026 02:16
URII:	http://shdl.mmu.edu.my/id/eprint/15712

Downloads

Edit (login required)