TamAko783/TPDD_Honglin_CLS-yolo26n-cls

yolo26n-cls — Tobacco Leaf Abnormality — Model Family

A family of YOLO26-nano classification models (ImageNet-pretrained, fine-tuned) for tobacco leaf abnormality identification on the TLA dataset. Two members:

Member	Weights	Training data
Base	base/yolo26n-cls-base.pt	canonical real images + Tier-3 lesion-paste + colour-safe oversampling (dataset_yolo_cls)
Synthetic	synth/yolo26n-cls-synth.pt	Base, fine-tuned on combined_synth = canonical + 762 SD-turbo diffusion rare-class variants

Both classify all 16 TLA classes. val/test are 100% real and identical for both.

Dataset credit

Data collected & annotated by Hong Lin, Rita Tse, Su-Kit Tang, Zhenping Qiang, Giovanni Pau et al. (Macao Polytechnic University) — TLA / TPDD. Original repository: https://github.com/honglin1226/Tobacco-Plant-Disease-Dataset. Cite TPDD (ICDIP 2022, 10.1117/12.2644288) and FREN (Frontiers in Plant Science 2024, 10.3389/fpls.2024.1333236). Models trained by TamAko783 (not the data creator). The source dataset carries no explicit license → these models are released for non-commercial research only.

Evaluation scope — COMMON classes only

Under the leakage-safe 80/10/10 split, the four long-tail classes (anthracnose, black_shank, tswv, genetic_abnormality) have only n=1 held-out image each, so a per-class score there is a coin-flip. They are predicted by the models but NOT benchmarked. All headline metrics below are over the 12 common classes (≥8 test images): wildfire, brown_spot, frog_eye, target_spot, tmv, cmv, pvy, weather_fleck, sunscald, potato_tuber_moth, nematodes, healthy.

Results (12 common classes, held-out real test = 120 images)

Metric	Base	Synthetic
Top-1	0.891	0.920
Macro-F1 (single split)	0.884	0.900
Macro-F1 (5-fold CV)	0.933 ± 0.012	—
Reference: FREN 16-way 10-shot	0.818	0.818

The 5-fold cross-validation (every real leaf rotated through the held-out fold) is the trustworthy benchmark for the Base model: macro-F1 0.933, top-1 ~0.936, tight across folds — comfortably past the paper's 0.818, with no single-image luck.

Per-class F1 (common 12, single-split test)

class	Base	Synthetic	Δ
wildfire	0.973	1.000	+0.03
weather_fleck	0.957	1.000	+0.04
tmv	0.889	0.963	+0.07
target_spot	0.842	0.952	+0.11
sunscald	0.857	0.929	+0.07
potato_tuber_moth	0.889	0.947	+0.06
healthy	0.909	0.952	+0.04
nematodes	0.750	0.800	+0.05
frog_eye	0.952	0.952	0
brown_spot	0.952	0.952	0
cmv	0.870	0.800	−0.07
pvy	0.769	0.545	−0.22

Takeaway: the Synthetic fine-tune raises overall common-class accuracy (top-1 +2.9pp, macro-F1 +1.6pp) and lifts most classes, but degrades the mosaic-virus classes (pvy, cmv) — the diffusion rare-class variants appear to blur virus-specific cues. Use Base when virus discrimination matters; use Synthetic for best overall top-1.

Usage (apply the training preprocessing!)

Both models were trained on Shades-of-Gray colour-constancy–normalised images. A raw predict("leaf.jpg") on an unnormalised photo is off-distribution. Use the repo's predict.py, or replicate it:

import numpy as np
from PIL import Image, ImageOps
from ultralytics import YOLO

def shades_of_gray(img, p=6):
    x = img.astype(np.float32)
    illum = np.maximum(np.power(np.mean(np.power(x, p), axis=(0, 1)), 1.0 / p), 1e-6)
    return np.clip(x * (illum.mean() / illum), 0, 255).astype(np.uint8)

model = YOLO("base/yolo26n-cls-base.pt")          # or synth/yolo26n-cls-synth.pt
im = ImageOps.exif_transpose(Image.open("leaf.jpg")).convert("RGB")
im = Image.fromarray(shades_of_gray(np.asarray(im)))
r = model.predict(im)
print(r[0].probs.top1, model.names[r[0].probs.top1])

Training (reproducible)

• yolo26n-cls.pt (ImageNet), imgsz 256, batch 64, optimizer auto, cos_lr, dropout 0.1, label_smoothing 0.1, Tier-1 colour-safe online augmentation.
• Base: 200 epochs (early-stopped ~28) on dataset_yolo_cls (1295 train: 1122 real + 74 lesion-paste + 99 oversample; val 138 / test 138, 100% real).
• Synthetic: Base fine-tuned 80 epochs on combined_synth (Base train + 762 diffusion variants of the rare classes; val/test unchanged, 100% real).
• Reproduce: scripts/00_build_all.py → train_part2.py (Base) → 11_synthetic_variants.py + 08b_combine_synth.py → train_part2.py --model base.pt (Synthetic). Benchmark: eval_kfold.py (k-fold) / eval_part2.py (single split).

Files

base/yolo26n-cls-base.pt, synth/yolo26n-cls-synth.pt, base/per_class_metrics.csv, synth/per_class_metrics.csv, this card.

License

Derived from the TLA/TPDD images, which carry no explicit license (https://github.com/honglin1226/Tobacco-Plant-Disease-Dataset). Released for non-commercial academic research only, with attribution and no ownership claim. Original authors may open an issue for takedown.