Advancing Image Classification through Parameter-Efficient Fine-Tuning: A Study on LoRA with Plant Disease Detection Datasets

Deeksha Aggarwal, Yash Mittal, Uttam Kumar

Keywords: Low-Rank Approximation, Parameter-Efficient Fine-Tuning, Image Classification, Plant Disease Detection

TL;DR: This study focuses on the application of LoRA, a low-rank approximation technique, within the realm of image classification for plant disease detection and evaluates its performance compared to traditional fine-tuning methods.

Abstract: Low-Rank Approximation (LoRA) has demonstrated remarkable efficiency in Large Language Models (LLMs), enabling the attainment of state-of-the-art results across various Natural Language Processing (NLP) tasks. This study shifts the focus to the application of LoRA, a low-rank approximation technique, within the realm of image classification for plant disease detection. The experiments unveil a notable reduction in trainable parameters to less than 1%. Notably, LoRA surpasses traditional fine-tuning methods, achieving a state-of-the-art accuracy of 73.83% on the label-deficient Pdoc dataset, 99.89% on the PlantVillage dataset and 81.81% on FieldPlant dataset. This research provides valuable insights at the intersection of PEFT and real-world plant disease detection, establishing a foundation for future explorations across diverse domains.

Submission Number: 204