Rice Leaf Disease Classification using Knowledge Distillation

Abstract

Rice is essential for global food security, particularly in Asian countries where it is a staple food and a significant contributor to the economy. Rice cultivation faces numerous challenges, including a vari-ety of illnesses that can significantly impact the yield and quality of the harvest. Conventional disease detection methods are slow, require a lot of manual effort, and are prone to mistakes, emphasizing the crit-ical requirement for automated approaches. This research explores the implementation of knowledge distillation, a method of transferring ex-pertise from advanced Models to more basic ones, for identifying diseases in rice. We condensed the expertise of complex Models by training them extensively, creating a streamlined and effective Model. We employed DenseNet121 as the Teacher Model, and the Student Model is equipped with dual pipeline Models to address overfitting. Both of these Mod-els passed through Quantization-Aware Training (QAT). The Student-Teacher collaborative Model achieved a precision of 98.47%, using sig-nificantly fewer parameters (5.9 million) than the Teacher-only Model.
Finally, the whole Model was quantized to an 8-bit integer from 32-bit float points. This leads to a fairer method for effectively and accurately isolating small and complex characteristics. This small-scale is ideal for use on portable gadgets in environments with restricted resources, en-abling farmers to quickly identify diseases for better crop management and increased worldwide food security measures.