jaganadhg/cheque-field-annotations

IDRBT Cheque Field Annotations

Bounding-box annotations for six standard fields in Indian bank cheques, derived from the publicly released IDRBT Cheque Image Dataset.

Images are not included. This dataset contains annotations only. The original TIFF images must be obtained directly from IDRBT under their terms of use. Filenames in this dataset correspond 1-to-1 with the images in the IDRBT release.

---

Dataset Summary

Property	Value
Total annotated cheques	112
Splits	Train 90 / Validation 11 / Test 11
Fields per cheque	6 (fixed)
Annotation format	Bounding box [xmin, ymin, xmax, ymax] (absolute pixels)
Original image format	TIFF, RGB, ~2365 × 1087 px
Original annotation format	Pascal VOC XML

---

Fields

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

Field	Key	Description
Date	date	Cheque date (typically 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	Cheque identifier (filename without extension)
filename	string	Original TIFF filename (e.g. Cheque 083654.tif)
image_width	int32	Original image width in pixels
image_height	int32	Original image height in pixels
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 in absolute pixels relative to the original image.

Data Splits

Split	Examples
Train	90
Validation	11
Test	11
Total	112

Splits are reproducible (random seed 42).

---

Usage

from datasets import load_dataset

dataset = load_dataset("jaganadhg/cheque-field-annotations")
print(dataset)
# DatasetDict({
#     train:      Dataset({features: [...], num_rows: 90}),
#     validation: Dataset({features: [...], num_rows: 11}),
#     test:       Dataset({features: [...], num_rows: 11})
# })

sample = dataset["train"][0]
print(sample["filename"])          # 'Cheque 083654.tif'
print(sample["image_width"])       # 2372
print(sample["date"])              # {'xmin': 1762, 'ymin': 65, 'xmax': 2329, 'ymax': 186}

Convert to COCO format

from datasets import load_dataset

FIELD_NAMES = ["date", "amount", "ifsc", "acno", "sign", "name"]
LABEL2ID    = {f: i + 1 for i, f in enumerate(FIELD_NAMES)}  # 1-indexed

dataset = load_dataset("jaganadhg/cheque-field-annotations")

def to_coco_row(example):
    """Convert one dataset row to a COCO-style annotation list."""
    annotations = []
    for field in FIELD_NAMES:
        bb = example[field]
        w  = bb["xmax"] - bb["xmin"]
        h  = bb["ymax"] - bb["ymin"]
        annotations.append({
            "category_id": LABEL2ID[field],
            "category":    field,
            "bbox":        [bb["xmin"], bb["ymin"], w, h],   # COCO: [x, y, w, h]
            "area":        w * h,
            "iscrowd":     0,
        })
    example["annotations"] = annotations
    return example

coco_dataset = dataset.map(to_coco_row)

Normalise coordinates for model training

def normalise(example):
    """Normalise boxes to [0,1] relative to image dimensions."""
    W, H = example["image_width"], example["image_height"]
    for field in ["date", "amount", "ifsc", "acno", "sign", "name"]:
        bb = example[field]
        example[f"{field}_norm"] = {
            "xmin": bb["xmin"] / W,
            "ymin": bb["ymin"] / H,
            "xmax": bb["xmax"] / W,
            "ymax": bb["ymax"] / H,
        }
    return example

normalised = dataset.map(normalise)

Use with IDRBT images (after downloading)

import os
from datasets import load_dataset
from PIL import Image

IMAGE_DIR = "/path/to/IDRBT_Cheque_Image_Dataset/300"
dataset   = load_dataset("jaganadhg/cheque-field-annotations")

def add_image(example):
    img_path = os.path.join(IMAGE_DIR, example["filename"])
    example["image"] = Image.open(img_path).convert("RGB")
    return example

dataset_with_images = dataset.map(add_image)

---

Field Layout

The spatial distribution of fields across cheques (normalised coordinates, mean ± std over all 112 cheques):

Field	Centre-x	Centre-y	Width	Height
date	0.85 ± 0.01	0.13 ± 0.01	0.26 ± 0.02	0.14 ± 0.01
amount	0.85 ± 0.01	0.42 ± 0.01	0.27 ± 0.01	0.14 ± 0.01
ifsc	0.22 ± 0.01	0.16 ± 0.01	0.13 ± 0.01	0.05 ± 0.01
acno	0.21 ± 0.05	0.53 ± 0.02	0.32 ± 0.05	0.11 ± 0.02
sign	0.90 ± 0.02	0.72 ± 0.04	0.14 ± 0.02	0.22 ± 0.04
name	0.51 ± 0.01	0.25 ± 0.01	0.96 ± 0.01	0.11 ± 0.01

Cheques follow a consistent layout: name is a full-width band near the top, date and amount are stacked on the right, ifsc and acno span the centre-left, and sign sits in the bottom-right.

---

Benchmark Results

A ResNet-50 regression model trained on the same annotations achieves the following on a held-out test set of 10 images. (The model was trained from an earlier HDF5 consolidation of these annotations in which 7 rows were corrupted and excluded — 105 usable images, 85/10/10 split — so its splits differ slightly from the 90/11/11 splits of this release.)

Field	IoU	Acc@0.5
date	0.528	50%
amount	0.572	80%
ifsc	0.506	60%
acno	0.579	90%
sign	0.437	30%
name	0.658	90%
Mean	0.547	67%

These are indicative single-run numbers, not a strong benchmark. A no-learning baseline that predicts each field's mean training box already reaches 0.691 mIoU / 80% accuracy on this task, so cheque layout is highly regular and absolute IoU should be read with that prior in mind. See the accompanying paper for the full controlled analysis (including a negative result on synthetic-data augmentation).

Pre-trained model: jaganadhg/cheque-field-regressor

---

Source Data

Original dataset published by the Institute for Development and Research in Banking Technology (IDRBT), Hyderabad, India.

• Original release: IDRBT Cheque Image Dataset
• Original URL: https://www.idrbt.ac.in
• Original format: TIFF images + Pascal VOC XML annotations

Annotations were converted from Pascal VOC XML to Parquet format for this release. No image data is included.

---

License

The annotations in this dataset are derived from the IDRBT Cheque Image Dataset. Please refer to IDRBT's terms of use before using this dataset for commercial purposes.

---

Citation

If you use this dataset in your research, please cite:

@dataset{idrbt-cheque-annotations-2026,
  title        = {IDRBT Cheque Field Annotations},
  author       = {Gopinadhan, Jaganadh},
  year         = {2026},
  publisher    = {HuggingFace},
  url          = {https://huggingface.co/datasets/jaganadhg/cheque-field-annotations},
  note         = {Annotations derived from the IDRBT Cheque Image Dataset}
}

For the original IDRBT dataset, please also cite:

@misc{idrbt-cheque-dataset,
  title        = {IDRBT Cheque Image Dataset},
  author       = {{Institute for Development and Research in Banking Technology}},
  howpublished = {\url{https://www.idrbt.ac.in}},
  year         = {2020}
}