---
license: cc-by-4.0
language:
- en
metrics:
- accuracy
- recall
pipeline_tag: image-to-text
tags:
- agriculture
- leaf
- disease
datasets:
- enalis/LeafNet
library_name: transformers
---

# 🌿 SCOLD: A Vision-Language Foundation Model for Leaf Disease Identification

**SCOLD** (Leaf Disases Vision-Language) is a multimodal model that maps **images** and **text descriptions** into a shared embedding space. It combines a [Swin Transformer](https://huggingface.co/timm/swin_tiny_patch4_window7_224) as the **image encoder** and [RoBERTa](https://huggingface.co/roberta-base) as the **text encoder**, projected to a 512-dimensional common space.

This model is developed for **cross-modal retrieval**, **few-shot classification**, and **explainable AI in agriculture**, especially for plant disease diagnosis from both images and domain-specific text prompts.

---

## 🚀 Model Details

| Component        | Architecture                             |
|------------------|-------------------------------------------|
| Image Encoder    | Swin Base (patch4, window7, 224 resolution) |
| Text Encoder     | RoBERTa-base                              |
| Projection Head  | Linear layer (to 512-D space)             |
| Normalization    | L2 on both embeddings                     |
| Training Task    | Contrastive learning                      |

The final embeddings from image and text encoders are aligned using cosine similarity.

---

### ✅ Intended Use
- Vision-language embedding for classification or retrieval tasks
- Few-shot learning in agricultural or medical datasets
- Multimodal interpretability or zero-shot transfer
---

## 🧪 How to Use

```python
import torch
from transformers import RobertaTokenizer
from torchvision import transforms
from PIL import Image
from modeling_lvl import LVL  # Replace with your module or package

# Load model
model = LVL()
model.load_state_dict(torch.load("pytorch_model.bin", map_location="cpu"))
model.eval()

# Text preprocessing
tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
text = "A maize leaf with bacterial blight"
inputs = tokenizer(text, return_tensors="pt")

# Image preprocessing
image = Image.open("path_to_leaf.jpg").convert("RGB")
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor()
])
image_tensor = transform(image).unsqueeze(0)

# Inference
with torch.no_grad():
    image_emb, text_emb = model(image_tensor, inputs["input_ids"], inputs["attention_mask"])
    similarity = torch.nn.functional.cosine_similarity(image_emb, text_emb)
    print(f"Similarity score: {similarity.item():.4f}")
```
Please cite this paper if this code is useful for you!

```
@misc{quoc2025visionlanguage,
  author       = {Quoc, K. N. and Thu, L. L. T. and Quach, L. D.},
  title        = {A Vision-Language Foundation Model for Leaf Disease Identification},
  year         = {2025},
  publisher    = {Authorea Preprints},
  url          = {10.36227/techrxiv.174062971.11176782/v1}
}
```
Demo in [here](https://leafclip.streamlit.app/)