---
license: mit
library_name: rfdetr
tags:
  - plant-disease
  - disease-detection
  - agriculture
  - computer-vision
  - object-detection
  - rf-detr
  - precision-agriculture
  - crop-health
datasets:
  - plant-disease-faxnj
metrics:
  - map
pipeline_tag: object-detection
---

# CropScan - Plant Disease Detection Model

CropScan is a plant disease detection model based on RF-DETR, designed to help farmers quickly identify health issues in their crops.

## Why CropScan?

Farming is hard work. Farmers face countless daily challenges: unpredictable weather, economic pressures, and most critically, crop diseases that can devastate entire harvests in just a few days.

**CropScan was built to:**

- **Help farmers** detect diseases early, before they spread
- **Reduce crop losses** through rapid and targeted intervention
- **Optimize treatment usage** by precisely identifying affected areas
- **Democratize access** to advanced diagnostic tools, once reserved for experts

Whether you're a small-scale farmer or a large producer, CropScan gives you the power to protect your crops with artificial intelligence.

## Detection Example

| Original Image | Detection Result |
|:--------------:|:----------------:|
| ![Original](examples/original.png) | ![Detection](examples/detection_result.png) |

The left image shows a leaf with disease symptoms. The right image shows CropScan's result: each diseased region is identified and segmented with precision using SAM2 integration.

## Technical Details

| Specification | Value |
|--------------|-------|
| **Architecture** | RF-DETR (medium) |
| **Task** | Object Detection / Disease Localization |
| **Performance** | mAP@50: 0.502 |
| **Model Size** | 134 MB |
| **Format** | PyTorch (.pth) |

## Usage

### Installation

```bash
pip install rfdetr torch torchvision
```

### Inference

```python
import torch
from rfdetr import RFDETRBase
from PIL import Image

# Load the model
model = RFDETRBase()
checkpoint = torch.load("checkpoint_best_total.pth", map_location="cpu")
model.load_state_dict(checkpoint)
model.eval()

# Load an image
image = Image.open("your_image.jpg")

# Run detection
with torch.no_grad():
    predictions = model(image)

# predictions contains bounding boxes of diseased regions
```

### SAM2 Integration (Recommended)

For precise segmentation masks instead of bounding boxes, combine CropScan with SAM2:

```python
from sam2.sam2_image_predictor import SAM2ImagePredictor

# Use CropScan boxes as prompts for SAM2
predictor = SAM2ImagePredictor.from_pretrained("facebook/sam2-hiera-small")
predictor.set_image(image)

for box in predictions.boxes:
    masks, scores, _ = predictor.predict(box=box, multimask_output=False)
    # masks contains the precise segmentation mask
```

## Training Data

This model was trained on the Plant Disease dataset from Roboflow Universe, containing images of leaves with various diseases.

```bibtex
@misc{plant-disease-faxnj_dataset,
    title = { Plant disease Dataset },
    type = { Open Source Dataset },
    author = { Project },
    howpublished = { \url{ https://universe.roboflow.com/project-oklwn/plant-disease-faxnj } },
    url = { https://universe.roboflow.com/project-oklwn/plant-disease-faxnj },
    journal = { Roboflow Universe },
    publisher = { Roboflow },
    year = { 2024 },
    month = { feb },
}
```

## Use Cases

- **Precision Agriculture**: Automated crop monitoring via drone or fixed camera
- **Field Diagnosis**: Mobile app for rapid disease identification
- **Agricultural Research**: Study of plant disease propagation
- **Education**: Teaching tool for agronomy students

## Limitations

- Trained primarily on PlantVillage-style images
- Best performance on individual leaf images with clear backgrounds
- SAM2 recommended for precise segmentation masks
- Does not replace expert agronomist diagnosis

## License

This model is distributed under the MIT license. You are free to use, modify, and distribute it for commercial or non-commercial purposes.

---

*Built with passion to support those who feed us.*