dganoun/leaf-classifier-resnext101

Leaf margin classifier (ResNeXt-101)

Classifies a leaf photo by the shape of its margin: dente (toothed edge) or lisse (smooth edge).

• Code: https://github.com/dihiaganoun/Leaf-Classification
• Live demo: https://leaf-margin-classifier.streamlit.app

Model

• Architecture: ResNeXt-101 32×8d (torchvision), ImageNet weights, new 2-class fc head
• Fine-tuning: layer3, layer4 and fc trained, earlier layers frozen
• Input: 224×224 RGB, ImageNet normalization
• Classes (output order): ["dente", "lisse"]
• File: resnext101_leaves.pth is a PyTorch state_dict

Results

Held-out test split (15% of ~6,500 images, 984 images):

Metric	Score
Accuracy	98.68%
F1	0.9861
Precision	0.9935
Recall	0.9787
Balanced accuracy	0.9864

Usage

import torch, torch.nn as nn
from huggingface_hub import hf_hub_download
from PIL import Image
from torchvision import models, transforms

path = hf_hub_download("dganoun/leaf-classifier-resnext101", "resnext101_leaves.pth")
model = models.resnext101_32x8d(weights=None)
model.fc = nn.Linear(model.fc.in_features, 2)
model.load_state_dict(torch.load(path, map_location="cpu", weights_only=True))
model.eval()

transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])
x = transform(Image.open("leaf.jpg").convert("RGB")).unsqueeze(0)
probs = model(x).softmax(1)[0]
print(dict(zip(["dente", "lisse"], probs.tolist())))

Or install the project package from GitHub and call leaf_classifier.load_model().

Limitations

• Trained on images of single leaves; accuracy on real-world photos with cluttered backgrounds has not been measured.
• See the GitHub README for notes on the training notebook.