wqm-zyl/wqm-zzl-swin-tiny-tomato-disease

🍅 Swin Tiny for Tomato Disease Classification (10 Classes)

Fine-tuned Swin Tiny model for tomato disease classification based on the PlantVillage dataset.

Language / 语言

• 🇺🇸 English
• 🇨🇳 中文

---

中文

模型简介

本模型基于 microsoft/swin-tiny-patch4-window7-224 进行全参数微调（Full Fine-tuning），用于番茄叶片病虫害分类，共支持 10 个类别。

模型基于 PlantVillage Tomato Disease Dataset 训练，并采用严格的数据集划分，最终结果使用完全未参与训练的独立测试集进行评测，确保 Zero Data Leakage（零数据泄漏）。

---

模型信息

项目	内容
基座模型	microsoft/swin-tiny-patch4-window7-224
模型结构	Swin Transformer Tiny
参数量	28M
输入尺寸	224 × 224
输出类别	10
微调方式	Full Fine-tuning

---

数据集

PlantVillage Tomato Disease Dataset

数据共 16,011 张图像。

数据划分	数量	说明
Training Set	800	模型训练（每类80张）
Validation Set	200	训练过程验证（每类20张）
Independent Test Set	15,011	最终评测（完全未参与训练）

最终 97.4% Accuracy 仅在 15,011 张独立测试集 上获得。

---

训练配置

项目	配置
Optimizer	AdamW
Learning Rate	1e-4
Scheduler	CosineAnnealingLR
Epochs	20

---

模型性能

Metric	Value
Overall Accuracy	97.4%
Lowest Class Accuracy	Early Blight (94.3%)
Highest Class Accuracy	Tomato Mosaic Virus (100%)

---

支持类别

ID	English
0	Bacterial Spot
1	Early Blight
2	Late Blight
3	Leaf Mold
4	Septoria Leaf Spot
5	Spider Mites
6	Target Spot
7	Tomato Yellow Leaf Curl Virus
8	Tomato Mosaic Virus
9	Healthy

---

使用方法

from transformers import AutoImageProcessor
from transformers import AutoModelForImageClassification
from PIL import Image

processor = AutoImageProcessor.from_pretrained(
    "wqm-zyl/wqm-zzl-swin-tiny-tomato-disease"
)

model = AutoModelForImageClassification.from_pretrained(
    "wqm-zyl/wqm-zzl-swin-tiny-tomato-disease"
)

image = Image.open("tomato_leaf.jpg").convert("RGB")

inputs = processor(images=image, return_tensors="pt")

outputs = model(**inputs)

predicted_id = outputs.logits.argmax(-1).item()

predicted_label = model.config.id2label[predicted_id]

confidence = outputs.logits.softmax(-1)[0, predicted_id].item()

print(predicted_label, confidence)

---

局限性

模型训练数据来自 PlantVillage 实验室环境。

在真实田间环境中（复杂背景、自然光照、手机拍摄等）准确率可能下降。

建议使用真实农业场景数据继续微调后再部署。

---

English

Model Overview

This repository provides a fully fine-tuned Swin Tiny model for Tomato Disease Classification.

The model is built upon microsoft/swin-tiny-patch4-window7-224 and supports 10 tomato disease categories.

The reported results are evaluated on an independent held-out test set with zero data leakage.

---

Dataset

The dataset contains 16,011 images.

Split	Images	Purpose
Training	800	Model training
Validation	200	Validation during training
Independent Test	15,011	Final evaluation

The reported 97.4% accuracy is obtained exclusively on the independent test set.

---

Training Configuration

• Base Model: microsoft/swin-tiny-patch4-window7-224
• Fine-tuning: Full Fine-tuning
• Parameters: 28M
• Optimizer: AdamW
• Learning Rate: 1e-4
• Scheduler: CosineAnnealingLR
• Epochs: 20

---

Performance

• Overall Accuracy: 97.4%
• Lowest Class: Early Blight (94.3%)
• Highest Class: Tomato Mosaic Virus (100%)

---

Usage

See the Python example above.

---

Limitations

The model is trained on PlantVillage laboratory images.

Performance may degrade on real-world images with complex backgrounds, natural lighting or mobile phone photography.

Further fine-tuning on field-collected data is recommended before deployment.

---

Authors

This project was completed collaboratively by:

• Wang Qimeng
  Hugging Face: https://huggingface.co/mengjoyful

• Zhu Zilong
  Hugging Face: https://huggingface.co/yanjuhuahuo

---

Contributions

Wang Qimeng

• Model training
• Full fine-tuning
• Hyperparameter optimization
• Hugging Face deployment
• Documentation

Zhu Zilong

• Dataset preparation
• Data preprocessing
• Model evaluation
• Experimental validation
• Documentation

---

Citation

@misc{swin-tiny-tomato-disease,
  author = {Wang Qimeng and Zhu Zilong},
  title = {Swin Tiny for Tomato Disease Classification},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/wqm-zyl/wqm-zzl-swin-tiny-tomato-disease}
}