# LMP-PM: a lightweight multi-path pruning method for plant leaf disease recognition

## TL;DR

This paper introduces LMP-PM, a new method to create efficient and accurate models for identifying plant leaf diseases, suitable for use in resource-limited agricultural settings.

## Contribution

The novel Lightweight Multi-Path Pruning Method (LMP-PM) enables flexible optimization of deep learning models for plant disease recognition.

## Key findings

- LMNet, optimized using LMP-PM, uses only 5.69% of the parameters and 3.80% of the FLOPs of the original model.
- LMNet achieves 99.23% accuracy on the Plant Village dataset and 87.27% on the AI 2018 Challenger dataset.
- LMP-PM improves classification accuracy while significantly reducing computational requirements.

## Abstract

Plant leaf diseases pose a significant threat to plant growth and productivity, necessitating accurate and timely identification. While high-performance deep learning models exist, their complexity often hinders deployment in real-world, resource-constrained agricultural settings. To address the need for efficient and accurate plant disease identification, we developed a novel lightweight approach named the Lightweight Multi-Path Pruning Method (LMP-PM). LMP-PM offers flexible lightweight optimization, configurable via pruning parameters and path expansion ratios, enabling users to balance significant reductions in model parameters and FLOPs against potential inference time increases, thereby tailoring model size, performance, and real-time needs to specific application scenarios. Specifically, we first constructed an original, high-performance, and complex model (OMNet) incorporating various structures and a three-branch parallel module (TBP block). We then applied LMP-PM to OMNet to perform lightweight processing, resulting in several lightweight models. Through extensive experimentation, we identified the optimal model that balances performance and complexity, which we named LMNet (Lightweight Multi-Path Network). LMNet demonstrates remarkable efficiency, utilizing only 5.69% of the parameters and 3.80% of the FLOPs of OMNet. Despite this substantial reduction in complexity, LMNet achieved superior accuracy: 99.23% on the Plant Village dataset, representing an improvement of 0.58% over OMNet, and 87.27% on the AI 2018 Challenger dataset, surpassing OMNet by 1.91%. These results highlight that LMP-PM successfully creates highly efficient models like LMNet, which not only drastically reduce computational resources but also improve classification accuracy. This flexibility and enhanced performance make LMNet particularly suitable for real-time plant disease identification in resource-constrained environments, offering a practical and effective solution for agricultural applications.