jaganadhg/cheque-synthetic-images

IDRBT Synthetic Cheque Images

A synthetic cheque image dataset generated from the IDRBT Cheque Image Dataset using a cut-paste augmentation technique. Each image is a novel composite: field content (date, amount, IFSC, account number, signature, payee name) is cropped from a real cheque and pasted onto a blank bank-specific canvas template. The original TIFF images are not distributed.

Images are fully synthetic. No original IDRBT cheque images are included. Field regions are copied from real cheques onto content-erased background templates, producing images that do not reproduce any complete original document.

---

Dataset Summary

Property	Value
Total synthetic cheques	295
Splits	Train 235 / Validation 30 / Test 30
Fields per cheque	6 (fixed)
Banks covered	Axis, Canara, ICICI, Syndicate
Annotation format	Bounding box [xmin, ymin, xmax, ymax] (absolute pixels)
Image format	PNG, RGB

---

Augmentation Technique

The method is inspired by Dwibedi et al. (2017) — Cut, Paste and Learn.

For each of the four bank types (Axis, Canara, ICICI, Syndicate), one blank canvas template was prepared by manually erasing all handwritten and printed field content from a representative cheque. At generation time:

1. A random cheque annotation is selected from the corresponding bank group.
2. Each of the six field regions is cropped from the matching original image.
3. The cropped patches are pasted onto the canvas template at the same absolute pixel coordinates.

This yields a new cheque image whose layout is exactly the canvas template but whose field content comes from a (randomly chosen) real cheque. The bounding-box annotations are carried over unchanged from the original.

Bank	Real cheques	Synthetic generated
Axis	87	79
Canara	10	91
ICICI	8	63
Syndicate	7	62
Total	112	295

337 images were attempted (87 Axis at a 1:1 ratio, 100/80/70 for the smaller banks at ~10:1 upsampling); 295 were generated. The difference is due to cheque numbers referenced in the bank metadata that have no corresponding TIFF in the source release.

---

Fields

Each cheque is annotated with exactly one bounding box per field:

Field	Key	Description
Date	date	Cheque date (top-right)
Amount (figures)	amount	Numeric amount (right column)
IFSC / branch code	ifsc	Bank branch identifier (mid-left)
Account number	acno	Full account number (centre)
Signature	sign	Handwritten signature region (bottom-right)
Payee name	name	"Pay to" name (full-width band)

---

Dataset Structure

Data Fields

Field	Type	Description
image_id	string	Synthetic cheque identifier
filename	string	PNG filename (e.g. axis_syn_0042.png)
bank	string	Bank type: axis, canara, icici, syndicate
image_width	int32	Image width in pixels
image_height	int32	Image height in pixels
image	Image	PNG-encoded cheque image
date	struct	{xmin, ymin, xmax, ymax}
amount	struct	{xmin, ymin, xmax, ymax}
ifsc	struct	{xmin, ymin, xmax, ymax}
acno	struct	{xmin, ymin, xmax, ymax}
sign	struct	{xmin, ymin, xmax, ymax}
name	struct	{xmin, ymin, xmax, ymax}

All coordinates are absolute pixels in the generated image (same coordinate space as the canvas template).

Data Splits

Split	Examples
Train	235
Validation	30
Test	30
Total	295

Splits are reproducible (random seed 42).

---

Usage

from datasets import load_dataset
from PIL import Image
import io

dataset = load_dataset("jaganadhg/cheque-synthetic-images")
print(dataset)

sample = dataset["train"][0]
print(sample["bank"])             # 'axis'
print(sample["image_width"])      # e.g. 2365
print(sample["date"])             # {'xmin': 1658, 'ymin': 78, 'xmax': 2325, 'ymax': 224}

# Access the image
img = sample["image"]             # PIL Image (loaded by HuggingFace automatically)
img.show()

Use for model training

from datasets import load_dataset
import torch
import torchvision.transforms.v2 as T

FIELD_NAMES = ["date", "amount", "ifsc", "acno", "sign", "name"]

dataset = load_dataset("jaganadhg/cheque-synthetic-images")

def make_target(example):
    W = example["image_width"]
    H = example["image_height"]
    boxes = []
    for field in FIELD_NAMES:
        bb = example[field]
        boxes.append([
            bb["xmin"] / W, bb["ymin"] / H,
            bb["xmax"] / W, bb["ymax"] / H,
        ])
    example["boxes_norm"] = boxes
    return example

dataset = dataset.map(make_target)

Combine with the annotation-only dataset

from datasets import load_dataset, concatenate_datasets

real_ann = load_dataset("jaganadhg/cheque-field-annotations")
synthetic = load_dataset("jaganadhg/cheque-synthetic-images")

# synthetic["train"] contains images + annotations
# Use synthetic for training, real annotations for evaluation

---

Source Data

Generated from the IDRBT Cheque Image Dataset, published by the Institute for Development and Research in Banking Technology (IDRBT), Hyderabad, India.

• Original URL: https://www.idrbt.ac.in
• Annotations source: jaganadhg/cheque-field-annotations

---

License

Apache 2.0.

The underlying field content originates from the IDRBT Cheque Image Dataset. Please refer to IDRBT's terms of use before using this dataset for commercial purposes.

---

Citation

@dataset{idrbt-cheque-synthetic-2026,
  title        = {IDRBT Synthetic Cheque Images},
  author       = {Gopinadhan, Jaganadh},
  year         = {2026},
  publisher    = {HuggingFace},
  url          = {https://huggingface.co/datasets/jaganadhg/cheque-synthetic-images},
  note         = {Synthetic images generated via cut-paste augmentation from the IDRBT Cheque Image Dataset}
}

Augmentation technique:

@inproceedings{dwibedi2017cutpaste,
  title     = {Cut, Paste and Learn: Surprisingly Easy Synthesis for Instance Detection},
  author    = {Dwibedi, Debidatta and Misra, Ishan and Hebert, Martial},
  booktitle = {ICCV},
  year      = {2017}
}