squirelmail/dataset-BotDetect-CAPTCHA-Generator

📦 CAPTCHA Datasets (style0–style59)

This repository contains CAPTCHA datasets for training CRNN+CTC models. Each archive dataset_*.tar.gz includes 60 styles (from BotDetect Captcha), structured as folders style0 through style59. Each style contains N images depending on the archive name.

---

🗂️ Available Archives

• dataset_500.tar.gz → 500 images per style (≈ 30,000 total)
• dataset_1000.tar.gz → 1,000 images per style (≈ 60,000 total)
• dataset_5000.tar.gz → 5,000 images per style (≈ 300,000 total)
• dataset_10000.tar.gz → 10,000 images per style (≈ 600,000 total)
• dataset_20000.tar.gz → 20,000 images per style (≈ 1,200,000 total)
• dataset_1000_rand.tar.gz → randomized variant with 1,000 images per style

Naming convention: dataset_{N}.tar.gz means each styleX folder holds exactly N PNG images.

---

📁 Directory Layout

/path/to/dataset
├── style0/
│   ├── A1B2C.png
│   ├── 9Z7QK.png
│   └── ...
├── style1/
│   ├── K9NO2.png
│   └── ...
└── ...
    └── style59/

• Filename = ground-truth label (5 uppercase alphanumeric chars), e.g. K9NO2.png.
• Image size = 50×250 pixels (H=50, W=250), grayscale PNG.
• Label rule = regex ^[A-Z0-9]{5}$ (exactly 5 chars, uppercase & digits).

---

🧰 Extraction

# example: extract into /workspace/dataset_1000
mkdir -p /workspace/dataset_1000
tar -xvzf dataset_1000.tar.gz -C /workspace/dataset_1000

---

✅ Quick File Counts

# total PNG files (depth 2 to only count inside style folders)
find /workspace/dataset_1000 -maxdepth 2 -type f -name '*.png' | wc -l

# per-style counts without a for-loop (prints "count styleX")
find /workspace/dataset_1000 -mindepth 2 -maxdepth 2 -type f -name '*.png' \
| awk -F/ '{print $(NF-2)}' | sort | uniq -c | sort -k2

Expected totals:

• dataset_500 → 500 × 60 = 30,000 files
• dataset_1000 → 60,000 files
• dataset_5000 → 300,000 files
• dataset_10000 → 600,000 files
• dataset_20000 → 1,200,000 files

---

🧪 Label Validation

# list filenames that violate the strict 5-char uppercase/digit rule
find /workspace/dataset_1000 -type f -name '*.png' \
| awk -F/ '{print $NF}' | sed 's/\.png$//' \
| grep -vE '^[A-Z0-9]{5}$' | head

CSV report via Python (pandas):

import os, re
import pandas as pd
from glob import glob

root = "/workspace/dataset_1000"
rows = []
for s in range(60):
    for p in glob(os.path.join(root, f"style{s}", "*.png")):
        rows.append({"style": f"style{s}", "filepath": p, "label": os.path.basename(p)[:-4]})

df = pd.DataFrame(rows)
bad = df[~df["label"].str.match(r"^[A-Z0-9]{5}$", na=True)]
print("Invalid labels:", len(bad))
if len(bad):
    bad.to_csv("invalid_labels.csv", index=False)

---

🧩 Example: Load to DataFrame

import os
from glob import glob
import pandas as pd

def load_dataset(root_dir):
    data = []
    for style_id in range(60):
        folder = os.path.join(root_dir, f"style{style_id}")
        for path in glob(os.path.join(folder, "*.png")):
            label = os.path.splitext(os.path.basename(path))[0]
            data.append((path, label, f"style{style_id}"))
    df = pd.DataFrame(data, columns=["filepath", "label", "style"])
    # enforce strict label rule
    df = df[df["label"].str.match(r"^[A-Z0-9]{5}$")]
    return df

df = load_dataset("/workspace/dataset_1000")
print(df.head(), len(df))

---

🔀 Merge Datasets (no loop)

Add new files without overwriting existing ones:

rsync -av \
  --ignore-existing \
  --include='style[0-5][0-9]/' \
  --include='style[0-5][0-9]/*.png' \
  --exclude='*' \
  /workspace/dataset_10000/  /workspace/dataset_20000/

Overwrite only if source is newer:

rsync -av --update \
  --include='style[0-5][0-9]/' \
  --include='style[0-5][0-9]/*.png' \
  --exclude='*' \
  /workspace/dataset_10000/  /workspace/dataset_20000/

---

🔒 Checksums

Optional: keep SHA256 for integrity.

sha256sum dataset_1000.tar.gz > dataset_1000.tar.gz.sha256
sha256sum -c dataset_1000.tar.gz.sha256

---

📝 Notes

• All images prepared for CRNN+CTC models with input (H, W) = (50, 250), grayscale.
• Character distribution: digits 0–9 and letters A–Z (uppercase).
• Each style emulates a distinct visual variant (font/noise/warp) from BotDetect.

---

📞 Contact

For questions, dataset issues, or custom subsets, please open an issue in this repository.