Skip to content

Latest commit

 

History

History
92 lines (68 loc) · 2.88 KB

README.md

File metadata and controls

92 lines (68 loc) · 2.88 KB

Light_asf_net: Fine-grained Disease Identification Network of Crop Leaves Based on Local Feature Enhancement

This repository contains the implementation of Light_asf_net, a lightweight dual-branch network for fine-grained classification of crop leaf diseases.

Abstract

Fine-grained classification of crop leaf diseases faces challenges such as large intra-class variation, inter-class ambiguity, and background interference. We propose Light_asf_net, which features:

  • Global Branch: ASF-former Transformer to capture leaf morphology.
  • Local Branch: MBConv module to extract lesion microstructures.
  • Adaptive Feature Fusion (AFF): Enables collaborative local-global enhancement.

Results:

  • Cassava Leaf Disease Dataset: 70.45% mAcc (37.8% improvement over ResNet-101) with only 6.03M parameters.
  • Rice Leaf Disease Dataset: 100% accuracy within 12 epochs.
  • Inference Speed: 35 FPS on mobile devices (Snapdragon 865).

Directory Structure

.
├── net/                # Network definitions (Light_asf_net, etc.)
├── models/             # ASF-former dependencies
├── tools/              # Utilities and Dataset classes
├── dataset_toolbox/    # Scripts for data processing
├── ASF-former-main/    # Upstream ASF-former reference
├── train.py            # Training script
├── predict.py          # Inference/Evaluation script
├── config.py           # Configuration file
└── requirements.txt    # Dependencies

Requirements

Install dependencies:

pip install -r requirements.txt

Key dependencies:

  • PyTorch >= 1.8.0
  • torchvision
  • albumentations
  • torchmetrics
  • tensorboard

Data Preparation

Rice Leaf Disease Dataset

  1. Place your images in Data/Rice Leaf Disease Images/<CategoryName>/.
  2. Use scripts in dataset_toolbox/ to generate the processed_data/ folder and train.csv.
    • Example: Run dataset_toolbox/3.dataset_division_new.py (you may need to adjust paths).

Cassava Leaf Disease Dataset

Similar structure applies. Ensure your CSV file format matches:

图片名,标签
image_name.jpg,label_index

Note: Update config.py with your DATA_DIR and LABEL_DIR paths.

Usage

Training

Configure parameters in config.py, then run:

python train.py

Inference / Evaluation

To evaluate trained models:

python predict.py

Configuration

Modify config.py to change:

  • DATA_DIR: Path to dataset images.
  • LABEL_DIR: Path to training labels CSV.
  • MODEL_NAME: "light_asf_net" (default).
  • BATCH_SIZE, LR, MAX_EPOCH: Training hyperparameters.

Reference

Research on Fine-grained Disease Identification Network of Crop Leaves Based on Local Feature Enhancement Liwei Fan, Siyan Liu, Kai Cai, Sixing Lu, Ke Zhang, Huixian Chen. School of Electronic Information and Control Engineering, Guangzhou University of Software.