README.md

metadata

license: mit
task_categories:
  - object-detection
tags:
  - image
  - computer-vision
  - document-analysis
  - financial-documents
  - table-structure
  - yolo
  - synthetic-data
pretty_name: Synthetic Bank Statement Table Structure Dataset
size_categories:
  - n<10k
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: image
      dtype: image
    - name: objects
      struct:
        - name: bbox
          list:
            list: float32
        - name: category
          list: int64
  splits:
    - name: train
      num_bytes: 1878958824.4
      num_examples: 8475
  download_size: 1816175365
  dataset_size: 1878958824.4

Synthetic Bank Statement Table Structure Dataset

A synthetically generated collection of bank statement images with pixel-perfect, automatically-produced bounding box annotations for table structure recognition (TSR).

🔑 In one sentence: fake bank statements + auto-generated YOLO labels for every table cell, built so you can train table-detection models (TATR, DETR, YOLO) without manual annotation.

---

At a Glance

Task	Object Detection → Table Structure Recognition
Format	PNG images + YOLO-format .txt labels
Classes	2 (0 = Header Cell, 1 = Body Cell)
Total images	8475
Splits	None yet — single unsplit set (train/val/test split not yet defined)
Source	100% synthetic, generated with Python + ReportLab
Real data?	❌ No real bank statements, customers, or financial records
License	MIT

---

Why This Dataset Exists

This dataset models clean, digitally-generated bank statement PDFs — the kind produced directly by a bank's own statement-generation system and downloaded as a native PDF (not a scanned paper document). Manually labeling table cells even in clean digital documents is slow and error-prone — humans miss edges, misjudge wrapped text boundaries, and disagree on where one cell ends and another begins. This dataset sidesteps that entirely: every bounding box is captured directly from the ReportLab drawing engine at the moment the PDF is rendered, so the label is the ground truth — there's no separate annotation step, and therefore no annotation drift or human error.

This makes it well-suited as pretraining or augmentation data for table structure models that will later be fine-tuned or validated on real bank statement PDFs.

---

What's Inside

Visual structure being detected

• 🔵 Header cells — column titles like Date, Description, Reference, Amount, Balance
• 🟢 Body cells — every individual transaction-data cell

Layout diversity

The generator doesn't produce one templated look — it varies:

• Portrait and landscape orientations
• Multiple banking templates and column counts
• Bordered, borderless, and zebra-striped tables
• Single- and multi-line (wrapped) transaction descriptions
• Dynamic row heights
• Multi-page statements with repeated headers across page breaks (each page is its own image/label pair — see Folder Structure for the naming convention)
• Randomized customer info, transaction narratives, and financial values

This diversity is the main value of the dataset — it's designed to expose a model to many table "shapes" rather than one clean layout repeated many times.

---

Folder Structure

structure/
├── images/
│   ├── anon_stmt_001_p1.png
│   ├── anon_stmt_001_p2.png
│   ├── anon_stmt_002_p1.png
│   └── ...
└── labels/
    ├── anon_stmt_001_p1.txt
    ├── anon_stmt_001_p2.txt
    ├── anon_stmt_002_p1.txt
    └── ...

Each image has a matching label file with the same base filename (e.g. images/anon_stmt_001_p1.png ↔ labels/anon_stmt_001_p1.txt).

Filename convention: anon_stmt_<statement_id>_p<page_number>

• anon_ — flags that the statement is fully anonymized/synthetic (no real customer data)
• stmt_<statement_id> — a unique ID per generated statement (e.g. 001, 002, ...)
• p<page_number> — page number within that statement (p1, p2, ...). Multi-page statements are stored as separate image/label pairs per page, not as a single multi-page file — so anon_stmt_001_p1.png and anon_stmt_001_p2.png belong to the same logical statement but are two independent training samples.

Label format

Standard YOLO format — one line per table cell:

class_id center_x center_y width height

• class_id: 0 (Header Cell) or 1 (Body Cell)
• All coordinates are normalized to image width/height (range 0–1)

Class ID	Label	Description
0	Header Cell	Column header cells (Date, Description, Reference, Amount, Balance, etc.)
1	Body Cell	Individual transaction-data cells — one box per non-header cell

---

Data Example

The table below shows two raw images from images/ next to the full contents of their matching label files from labels/ — no annotation overlay drawn on top, just what you'll actually find in the dataset folders. The two samples use different templates to illustrate the layout diversity described above.

Image (images/anon_stmt_001_p1.png)	Image (images/anon_stmt_002_p1.png)
Label (labels/anon_stmt_001_p1.txt)	Label (labels/anon_stmt_002_p1.txt)
0 0.083 0.042 0.150 0.018 0 0.245 0.042 0.150 0.018 0 0.407 0.042 0.150 0.018 0 0.569 0.042 0.150 0.018 0 0.731 0.042 0.150 0.018 1 0.083 0.071 0.150 0.022 1 0.245 0.071 0.150 0.022 1 0.407 0.071 0.150 0.022 1 0.569 0.071 0.150 0.022 1 0.731 0.071 0.150 0.022 1 0.083 0.099 0.150 0.022 1 0.245 0.099 0.150 0.022 ...	0 0.071 0.038 0.130 0.020 0 0.213 0.038 0.130 0.020 0 0.355 0.038 0.130 0.020 0 0.497 0.038 0.130 0.020 0 0.639 0.038 0.130 0.020 0 0.781 0.038 0.130 0.020 1 0.071 0.064 0.130 0.024 1 0.213 0.064 0.130 0.024 1 0.355 0.064 0.130 0.024 1 0.497 0.064 0.130 0.024 1 0.639 0.064 0.130 0.024 1 0.781 0.064 0.130 0.024 ...

Each line is class_id center_x center_y width height, normalized 0–1 against image dimensions — 0 for header cells, 1 for body cells. The number of lines in a label file equals the number of cells visible in its matching image.

---

Load with the Hugging Face Hub (download files directly)

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="Panhapich/bank-statement-structure-recognition",
    repo_type="dataset"
)
print(local_dir)

Use with a YOLO training pipeline (e.g. Ultralytics)

# data.yaml
path: ./structure
train: images   # currently single unsplit set — update once splits exist
val: images
nc: 2
names: ["header_cell", "body_cell"]

from ultralytics import YOLO

model = YOLO("yolov8n.pt")
model.train(data="data.yaml", epochs=50, imgsz=1024)

Convert YOLO labels → Table Transformer (TATR) / COCO-style boxes

from PIL import Image

def yolo_to_xyxy(line, img_w, img_h):
    class_id, cx, cy, w, h = map(float, line.split())
    x1 = (cx - w / 2) * img_w
    y1 = (cy - h / 2) * img_h
    x2 = (cx + w / 2) * img_w
    y2 = (cy + h / 2) * img_h
    return int(class_id), [x1, y1, x2, y2]

img = Image.open("images/anon_stmt_001_p1.png")
with open("labels/anon_stmt_001_p1.txt") as f:
    boxes = [yolo_to_xyxy(line, *img.size) for line in f if line.strip()]

---

Annotation Methodology

Unlike datasets where humans draw boxes after the fact, every annotation here is produced during document generation:

1. A synthetic bank statement is procedurally composed (random bank template, customer info, transactions).
2. ReportLab renders the table to a PDF/image.
3. As each cell is drawn, its exact pixel coordinates are captured at render time.
4. Coordinates are converted to normalized YOLO format and written alongside the image.

Result: pixel-perfect alignment with zero manual labeling and zero annotation drift — the kind of clean signal that's hard to get from human-labeled data, but also worth knowing about when judging how the dataset will generalize (see Limitations below).

---

Compatible Models

• microsoft/table-transformer-structure-recognition-v1.1-all
• Table Transformer (TATR)
• DETR-based table structure models
• YOLO-family object detectors (YOLOv5/v8/v9/v10, etc.)
• Other transformer-based document layout/understanding architectures

Intended Applications

• Table Structure Recognition (TSR)
• Document Layout Analysis
• Intelligent Document Processing (IDP)
• Financial document understanding & information extraction
• OCR pipeline development (cell localization prior to text extraction)
• Document AI benchmarking

---

Limitations & Intended Use

Being upfront about what this dataset is not, so it's used appropriately:

• Not modeling scanned documents. This dataset simulates clean, digitally-generated bank statement PDFs (the kind a bank produces natively), not scanned paper documents — there's no scanner skew, smudging, or photographic noise either way. If your real-world target documents are scanned/photographed paper statements, this dataset won't cover that domain gap, and you'd need separate scan-augmentation or real scanned samples.
• Synthetic template variety vs. real-world template variety. While the generator produces many layouts (bordered/borderless, zebra-striping, multi-page, etc.), it can't fully replicate the sheer diversity of real banks' actual statement designs, fonts, and branding. Models trained only on this data may need fine-tuning on a sample of real bank statement PDFs before deployment.
• No real financial data. All customer info, transaction narratives, and amounts are randomized/fabricated — this dataset contains no PII and no real institution branding.
• Single unsplit set. There is currently no official train/val/test split — if you need one, you'll need to partition it yourself (e.g. a random 80/10/10 split).
• Two classes only. This dataset labels header vs. body cells, not finer structure like row/column spans, merged cells, or currency-specific fields.

Best used for: pretraining, data augmentation, or benchmarking table-structure detectors before fine-tuning on a smaller set of real, manually verified bank statement PDFs.

---

Dataset Origin

This dataset is entirely synthetic. No real bank statements, customer information, financial records, proprietary templates, or institution-specific branding are included anywhere in the dataset. All documents were procedurally generated using Python and ReportLab for research purposes.

---

Citation

If you use this dataset in academic research, please cite your corresponding publication describing the dataset and generation methodology.

@dataset{synthetic_bank_statement_tsr,
  title        = {Synthetic Bank Statement Table Structure Dataset},
  author       = {Uk, Panhapich},
  year         = {2026},
  note         = {Synthetically generated for table structure recognition research}
}

License

Released under the MIT License.