TamAko783/TPDD_Honglin_CLS

Tobacco Leaf Abnormality (TLA) — 16-class YOLO classification split

A cleaned, relabeled, leakage-safe train/val/test split of the TLA (Tobacco Leaf Abnormality) dataset, packaged in the Ultralytics image-classification folder layout (train/val/test/<class>/) and targeting yolo26n-cls.

This card is generated from dataset_card.md and copied into the tree by 00_build_all.py; do not hand-edit the copy inside dataset_yolo_cls/.

⚠️ Dataset credit — original authors

The underlying images were collected and annotated by Hong Lin, Rita Tse, Su-Kit Tang, Zhenping Qiang, Giovanni Pau and colleagues at the Macao Polytechnic University. TLA is the 16-class refinement of their TPDD (Tobacco Plant Disease Dataset). Original dataset repository: https://github.com/honglin1226/Tobacco-Plant-Disease-Dataset

All credit for the data belongs to them. This repository only performs data processing (cleaning, relabeling, a leakage-safe split, and optional train-only augmentation) and redistributes it for reproducible benchmarking. If you use this data, please cite the original papers below.

Processing & repackaging by TamAko783. The uploader is not the creator of the images and claims no ownership of the original data.

Source papers (please cite)

• TPDD — Tobacco Plant Disease Dataset. Hong Lin, Rita Tse, Su-Kit Tang, et al. SPIE — International Conference on Digital Image Processing (ICDIP) 2022. DOI: 10.1117/12.2644288
• A few-shot learning method for tobacco abnormality identification (TLA / FREN). Hong Lin, Zhenping Qiang, Rita Tse, Su-Kit Tang, Giovanni Pau. Frontiers in Plant Science, 2024. DOI: 10.3389/fpls.2024.1333236 (PMC11055634) — defines the 16-class taxonomy and reports the few-shot benchmark (60.7% 16-way 1-shot → 81.8% 16-way 10-shot).
• Dataset repository: https://github.com/honglin1226/Tobacco-Plant-Disease-Dataset

What's inside

train/  <16 class folders>/ *.jpg   # real + *_paste* (Tier-3 lesion) + *_aug* (oversample)
val/    <16 class folders>/ *.jpg   # 100% real
test/   <16 class folders>/ *.jpg   # 100% real

All images are Shades-of-Gray colour-constancy normalised (illumination white-balance), applied identically to every split. When you run inference, apply the same white-balance or you will be off-distribution (see the model card's predict.py).

The training set is expanded with train-only synthetic images: Tier-3 lesion copy-paste (*_paste*) for the rare classes, and Tier-1 colour-safe oversampling (*_aug*) up to a 60-image floor. To train on real images only, skip files matching *_paste* and *_aug*. val/ and test/ are 100% real. Rotation is left to the trainer's online augmentation (degrees=30), not baked to disk.

The 16 classes (folder name = label):

#	class (slug)	super-category	#	class (slug)	super-category
1	wildfire	Bacteria	9	pvy	Virus
2	brown_spot	Airborne fungi	10	tswv	Virus
3	frog_eye	Airborne fungi	11	weather_fleck	Nonparasitic
4	anthracnose	Airborne fungi	12	sunscald	Nonparasitic
5	target_spot	Soil-borne fungi	13	genetic_abnormality	Nonparasitic
6	black_shank	Soil-borne fungi	14	potato_tuber_moth	Pest-trace
7	tmv	Virus	15	nematodes	Pest-trace
8	cmv	Virus	16	healthy	Others

Split counts

split	real	lesion-paste (_paste)	oversample (_aug)	total
train	1122	74	99	1295
val	138	0	0	138
test	138	0	0	138
total	1398	74	99	1571

Lesion copy-paste adds genuinely novel diseased leaves for the long tail (anthracnose +17, tswv +26, genetic_abnormality +26, pvy +5). Tier-1 oversampling then tops each under-floor class up to 60 train images. black_shank has no usable TV6 lesion crops (its TV6 folder was downscaled duplicates), so it relies on oversampling only — with just 9 real whole-leaf images total, treat it (and the other rare classes) with care.

⚠️ Rare-class evaluation. Under 80/10/10, anthracnose, black_shank, tswv and genetic_abnormality have only 1 val and 1 test image each, so a single-split per-class F1 is a coin-flip. For a trustworthy benchmark use the stratified group k-fold evaluation (scripts/eval_kfold.py), which rotates every real leaf through the held-out fold and reports mean ± std.

How it was processed

The raw TLA dataset ships as two sections — TV3 (whole-leaf, 696 imgs) and TV6 (disease-fragment crops, 734 imgs). Pipeline (scripts/00_build_all.py):

1. Clean — drop thumbs.db, Windows 副本 copies, and near-duplicates (DCT pHash, within-class).
2. Relabel — map raw folder numbers to the canonical 16-class taxonomy; TV6's 5 multi-symptom wildfire sub-folders collapse to wildfire.
3. Group-aware split (80/10/10) — the split unit is the source leaf (a leaf's whole-leaf image + its fragment crops move together), stratified by class with a fixed seed. Critical: TV6 crops are derived from TV3 leaves (hundreds of shared filename stems), so a naive per-image split would leak. A gate asserts no source leaf appears in more than one split.
4. Materialize + colour constancy — EXIF-transpose, convert RGB, apply Shades-of-Gray white balance (applied to all splits), re-save JPEG.
5. Augment (train only) — (a) Tier-3 lesion copy-paste: real TV6 lesion crops blended onto host leaves (*_paste*); (b) Tier-1 oversample: colour-safe geometric/photometric variants up to a 60-image floor (*_aug*). All synthetics live only in train/. val/ and test/ are 100% real.

⚠️ A small number of rare-class TV6 fragments were sequentially renamed in the original data, so their source leaf is unrecoverable; those are placed train-only to stay leakage-safe. Because such a fragment could be a crop of a TV3 leaf now in val/test for the same class, a small residual leakage risk remains for the rare classes — another reason to trust the k-fold estimate.

Usage (Ultralytics)

from ultralytics import YOLO
model = YOLO("yolo26n-cls.pt")
model.train(data="path/to/this/dataset", imgsz=256, epochs=200)

Citation

@inproceedings{lin2022tpdd,
  title     = {Tobacco Plant Disease Dataset (TPDD)},
  author    = {Lin, Hong and Tse, Rita and Tang, Su-Kit and others},
  booktitle = {International Conference on Digital Image Processing (ICDIP)},
  year      = {2022},
  doi       = {10.1117/12.2644288},
  note      = {Data: https://github.com/honglin1226/Tobacco-Plant-Disease-Dataset}
}

@article{lin2024tla,
  title   = {A few-shot learning method for tobacco abnormality identification},
  author  = {Lin, Hong and Qiang, Zhenping and Tse, Rita and Tang, Su-Kit and Pau, Giovanni},
  journal = {Frontiers in Plant Science},
  year    = {2024},
  doi     = {10.3389/fpls.2024.1333236}
}

License

The original TLA/TPDD images are © their authors (Macao Polytechnic University) and are distributed from https://github.com/honglin1226/Tobacco-Plant-Disease-Dataset, which states no explicit license. Because the source grants no redistribution rights, this repackaging is shared for non-commercial academic research and reproducible benchmarking only, on a good-faith fair-use basis, with full attribution and no claim of ownership. Consult the source papers/repository for terms and cite the original work. If you are an original author and have any concern about this redistribution, please open an issue and it will be taken down.