LeafGen AI - models
Plant-disease classifier (ResNet18, 38 classes, 14 crops) and a conditional DCGAN that generates synthetic leaf images (64x64). Built for a Generative-AI PBL project (SDG 2 - Zero Hunger). Dataset: PlantVillage (lab photos).
Research prototype. Trained on plain-background lab photos; not reliable on real field photos, and it only knows these 14 crops.
Files
| File | What it is |
|---|---|
cnn_real.pt |
ResNet18 trained on real data only (recommended, best overall) |
cnn_aug.pt |
Same CNN trained on real + GAN-synthetic images (experiment model) |
generator.pt |
Conditional DCGAN generator (noise 100 + one-hot class -> 64x64 RGB) |
classes.json |
The 38 class names (index = model output index) |
rare.json |
Indices of the 6 "rare" classes used in the experiment |
results.json, assets/*.png |
Experiment results, confusion matrices, real-vs-synthetic figure |
onnx/*.onnx |
ONNX versions (if present): input image float (N,3,64,64) values 0-255 -> output probs; generator inputs z (N,100), onehot (N,38) -> images 0-255 |
leafgen_model.py |
Ready-to-use loading + inference code |
Quick start
from leafgen_model import load_classifier, classify, load_generator, generate
clf = load_classifier() # recommended model
print(classify(clf, rgb_image)) # top-3 [(class, prob), ...]
gen = load_generator()
imgs = generate(gen, "Apple___Cedar_apple_rust", n=8)
Preprocessing matters: resize to 64x64 (INTER_AREA), then CLAHE (clip 2.0, tile 4x4) on the L channel in LAB colour space. Skipping this hurts accuracy.
Results (3 seeds, 8,146 real test images)
| Metric | A: real only | B: real + synthetic |
|---|---|---|
| Macro-F1 (38 classes) | 0.9784 | 0.9708 |
| Accuracy | 98.42% | 97.78% |
| Mean recall on 6 rare classes (100 real train images each) | 93.56% | 94.02% |
Synthetic images did not improve overall performance; the rare-class gain is within seed-to-seed noise.
Limitations
- PlantVillage = plain-background lab photos; one leaf appears in many photos, so random splits can inflate scores.
- 64x64 resolution, one GAN architecture, 3 seeds.
- Only 14 crops; any other plant will get a (wrong) answer from these 38 classes - use the confidence value (low = unsure).
Dataset: Hughes & Salathe (2015), PlantVillage. Check the dataset licence before redistributing images.