beans-vit-disease-classifier

Beans ViT Disease Classifier

ViT-B/16 fine-tuned to classify bean leaf images into three classes:

• angular_leaf_spot
• bean_rust
• healthy

Model

• Architecture: Torchvision ViT-B/16
• Initialization: ImageNet-1K pretrained weights
• Input: RGB image
• Evaluation preprocessing: resize to 256 pixels and center crop to 224 x 224
• Total parameters: 85,800,963
• Parameters trainable during fine-tuning: 14,179,587

Training occurred in two stages:

1. Training only the classification head.
2. Fine-tuning Transformer blocks 10 and 11, the final normalization layer, and the classification head.

The model was selected exclusively using validation macro-F1.

Results

Test set:

• Images: 128
• Correct predictions: 124
• Errors: 4

All healthy images were classified correctly. The four errors occurred between angular_leaf_spot and bean_rust.

Comparison with the custom CNN

The ViT obtained higher test performance, but it is approximately 71 times larger than the custom CNN. The CNN may remain preferable for resource-constrained deployment.

Usage

Download the repository and use inference.py.

Example:

from PIL import Image
from inference import load_model, predict

model, config, device = load_model(".")

image = Image.open("bean_leaf.jpg")

results = predict(
    image,
    model,
    config,
    device
)

print(results)

Limitations

• The test split contains only 128 images.
• The model was evaluated on a single dataset.
• Mild or visually ambiguous symptoms may cause confusion between the two diseases.
• Key diagnostic evidence such as leaf underside, raised pustules, and field context may not be visible.
• The comparison with the earlier CNN is descriptive and was performed after the test split had already been inspected.
• This model is an experimental image classifier and must not replace diagnosis by a qualified agricultural specialist.