Darayut/Textline-Detection-Dataset

---
license: cc-by-4.0
language:
- km
- en
task_categories:
- object-detection
tags:
- khmer
- text-detection
- ocr
- document-analysis
- yolo
- synthetic
- scene-text
pretty_name: Khmer Text Detection (Ultimate)
size_categories:
- 10K<n<100K
---

# Textline Detection — Ultimate Dataset

A large-scale, multi-source dataset for **textline detection** using YOLO-format bounding box annotations.
Combines real scene text, document layout images, and synthetically generated Khmer document images.

---

## Dataset Summary

| Split | Images |
|-------|--------|
| Train | 30,658 |
| Val | 2,764 |
| **Total** | **33,422** |

### Classes

| ID | Name | Description |
|----|------|-------------|
| 0 | `text_line` | A line of Khmer or mixed-script text |
| 1 | `image` | An embedded image/figure region within a document |

---

## Data Sources

### 1. Real Scene Text
Images captured in natural environments — street signs, storefronts, billboards, and handwritten Khmer documents.
Derived from the **DonkeySmall** base dataset (~26K images).

### 2. Document Layout
Images from the **DocLayNet** multi-class document layout corpus, re-labelled for the two-class schema.
Covers documents: official reports, newspapers, and books.

### 3. Synthetic Khmer Documents
Programmatically generated document images using custom Khmer text rendering pipelines.
Fonts, sizes, backgrounds, and layouts are randomised. Labels are auto-generated (zero annotation cost).

---

## Annotation Format

Labels follow **YOLO v8** format — coordinates normalised to `[0, 1]`:

```
<class_id> <cx> <cy> <width> <height>
```

The `annotations` field is a JSON-serialised list:

```json
[
{"class_id": 0, "cx": 0.512, "cy": 0.234, "w": 0.310, "h": 0.045},
{"class_id": 1, "cx": 0.720, "cy": 0.600, "w": 0.200, "h": 0.250}
]
```

---

## Dataset Fields

| Field | Type | Description |
|-------|------|-------------|
| `image` | `Image` | Decoded PIL image |
| `image_path` | `string` | Original file path at collection time |
| `source` | `string` | `real_scene_text` / `doclaynet_khmer` / `synthetic_khmer_doc` |
| `split` | `string` | `train` or `val` |
| `annotations` | `string` | JSON list of YOLO bounding boxes |
| `num_objects` | `int32` | Number of annotated objects |

---

## Usage

```python
from datasets import load_dataset

ds = load_dataset("Darayut/Multilingual-Textline-Detection-Dataset")

sample = ds["train"][0]
print(sample["source"], sample["num_objects"])
sample["image"].show()
```

### Convert back to YOLO label files

```python
import json
from pathlib import Path

def save_labels(split, out_dir):
Path(out_dir).mkdir(parents=True, exist_ok=True)
for row in ds[split]:
stem = Path(row["image_path"]).stem
anns = json.loads(row["annotations"])
with open(f"{out_dir}/{stem}.txt", "w") as f:
for a in anns:
f.write(f"{a['class_id']} {a['cx']:.6f} {a['cy']:.6f} {a['w']:.6f} {a['h']:.6f}\n")

save_labels("train", "labels/train")
save_labels("val", "labels/val")
```

### YAML config for YOLOv8 training

```yaml
train: /path/to/images/train
val: /path/to/images/val
nc: 2
names:
0: text_line
1: image
```

---

## Limitations

- Synthetic images may not capture all real-world degradation (blur, skew, lighting variation).
- Scene-text labels are semi-automatic and may have occasional missed detections.
- Dataset is primarily Khmer script; other scripts appear only incidentally.

---

## License

[CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) — free to use, share, and adapt with attribution.

---