jsdnrs/ICDAR2019-SROIE

ICDAR2019's Scanned Receipts OCR and Information Extraction (SROIE)

The ICDAR2019 SROIE dataset was originally published by Huang et al. for the 15th International Conference on Document Analysis and Recognition (ICDAR2019) Robust Reading Challenge on Scanned Receipts OCR and Information Extraction (SROIE).

This work presents an extension of the original ICDAR2019 SROIE dataset, including 14 receipt annotations missing from the original Task 3 test dataset, in a format integrated with Hugging Face's datasets. This dataset also includes corrections to the original data and image metadata to easily integrate with a diverse set of multimodal models with only minor post-processing.

The additions and changes in this dataset are documented in an associated Github repository: https://github.com/jsdnrs/ICDAR2019-SROIE

Dataset Summary

The ICDAR2019 SROIE dataset contains 987 receipt images in English, annotated with key fields (company, date, address, and total) and bounding boxes with transcripts for all text regions in the images.

The original dataset was created to provide a large-scale and well-annotated receipt dataset with the goal of recognizing text from scanned structured and semi-structured receipts, and invoices in general. This dataset also includes "some special features and challenges, e.g., some receipts having poor paper quality, poor ink and printing quality; low resolution scanner and scanning distortion; folded invoices; too many unneeded interfering texts in complex layouts; long texts and small font sizes" (Huang et al., 2021).

Supported Tasks and Leaderboards

• Key Information Extraction (KIE) - Includes annotations for the company, date, address, and total for each receipt image. Additional annotations can easily be added for other fields.

• Optical Character Recognition (OCR) - Includes bounding boxes and transcriptions (words, numbers, and/or symbols) of the text contained within those bounding boxes.

• Document Visual Question Answering (VQA) - Easily extended to support VQA tasks using the existing annotations, such as:

  • "What is the name of the company on this receipt?"
  • "What is the date shown on this receipt?"
  • "What is the address of the company on this receipt?"
  • "What is the final total cost shown on this receipt?"

This dataset is not known to be featured in any leaderboards as of this writing.

Languages

The dataset contains English-language receipts produced in Malaysia between 2016 and 2018.

Data Structure

Data Instances

Each data point includes a receipt image with associated fields for image_size, entities, words, and bboxes:

• bboxes and words include the bounding boxes and associated text transcripts for the OCR task (Tasks 1 and 2)
• entities includes the fields from the key information extraction task (Task

An example from the dataset is as follows (the words and bboxes arrays are truncated for brevity):

{
  "image": <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=463x605 at ...>,
  "image_size": {
    "width": 463,
    "height": 605
  },
  "entities": {
    "company": "ABC HO TRADING",
    "date": "09/01/2019",
    "address": "NO.2&4, JALAN HARMONI 3/2, TAMAN DESA HARMONI. 81100 JOHOR BAHRU JOHOR",
    "total": "31.00"
  },
  "words": [
    "TAN CHAY YEE",
    "ABC HO TRADING",
    ... 33 more elements ...
  ],
  "bboxes": [
    [143, 43, 362, 80],
    [129, 86, 309, 111],
    ... 33 more elements ...
  ]
}

The following figure shows the receipt image from the above example with the bounding boxes overlaid in red using torchvision:

Additional notes on the data instances:

• All text is capitalized even though the text in the receipt images may be mixed-case.
• The integer coordinates for the bounding boxes are not normalized to the image dimensions. Instead they match the default 'xyxy' format used by the torchvision utilities.
• There is a one-to-one mapping of the elements in the words and bboxes arrays where each element in the words array represents the words, numbers, and/or symbols within the corresponding bounding box coordinates from the bboxes array.

⚠️ Note: bboxes are in raw pixel values, when using LayoutLM they must be normalized using the following function (see Hugging Face's LayoutLM Notes):

def normalize_bbox(bbox, width, height):
  return [
    int(1000 * (bbox[0] / width)),
    int(1000 * (bbox[1] / height)),
    int(1000 * (bbox[2] / width)),
    int(1000 * (bbox[3] / height))]

Data Fields

• image: <PIL.JpegImagePlugin.JpegImageFile image mode=RGB ...>
• image_size: mapping of the width and height of the image (useful for post-processing, if needed)
• entities: mapping of the key fields (company, date, address, and total)
• words: array of text region strings containing one or more words, numbers, or symbols
• bboxes: array of integer pixel coordinates (unnormalized) of the text region's bounding box in [x0, y0, x1, y1] format, representing corners with (x0, y0) being top left and (x1, y1) being bottom right

Data Splits

The dataset is split into a 626 training and 347 test images, each with associated annotations. The criterion for splitting the data is not known.

Dataset Creation

Curation Rationale

As stated above, the original dataset was created to provide a large-scale and well-annotated receipt dataset with the goal of recognizing text from scanned structured and semi-structured receipts, and invoices in general.

Source Data

The source data is images of receipts or invoices, largely from Malaysia between 2016 and 2018. The annotations, including bounding boxes and key fields, are believed to be expert-generated.

Initial Data Collection and Normalization

The authors did not provide any details about the data collections process. However, they do mention that the dataset "has some special features and challenges, e.g., some receipts having poor paper quality, poor ink and printing quality; low resolution scanner and scanning distortion; folded invoices; too many unneeded interfering texts in complex layouts; long texts and small font sizes" (Huang et al., 2021).

The original bounding box format was given as rectangles with four vertices [x1, y1, x2, y2, x3, y3, x4, y4], which are in clockwise order starting from the top. This format is equivalent to the 'xyxyxyxy' format in torchvision. This data was converted by the 'xyxy' format in torchvision by storing only the top left and bottom right vertices [x1, y1, x3, y3].

Refer to the Modifications to the Dataset section of the associated Gihub repository for other additions and changes to the dataset.

Who are the source language producers?

Since the dataset includes images of receipts, it should be considered machine-generated by point-of-sale systems then printed and photographed/scanned. While the content of the receipts reflects human purchases, the receipt information and layout is semi-structured.

Annotations

Annotation process

The original publication describes the annotation format but does not provide details about the annotation workflow, guidelines, or validation process. Each receipt image was annotated with:

• Word/region-level text bounding boxes represented with four vertices in clockwise order
• A text transcript for each bounding box
• Key information fields (company, date, address, and total)

Who are the annotators?

The annotations are provided as part of the ICDAR 2019 Robust Reading Challenge on Scanned Receipts OCR and Information Extraction. However, the publication does not report demographic, professional, or institutional information about the specific annotators. It is also not specified whether the annotations were produced by academic researchers, industry practitioners, or hired annotators.

Personal and Sensitive Information

The original authors noted that "to address the potential privacy issue, some sensitive fields (such as name, address and contact number etc) of the receipts are blurred" (Huang et al., 2021). However, the dataset still includes many company names and addresses.

Considerations for Using the Data

Social Impact of Dataset

As the original authors have already stated in their publication, this dataset "plays critical roles in streamlining document-intensive processes and office automation in many financial, accounting and taxation areas" and "there is an urgent need to research and develop fast, efficient and robust SROIE systems to reduce and even eliminate manual work" (Huang et al., 2021).

Ultimately, this dataset supports research in automating document-heavy workflows such as receipt and invoice processing.

Discussion of Biases

As mentioned above, the receipts are largely from Malaysia between 2016 and 2018. Therefore, models trained on this dataset might over-index on specific currency symbols (RM), receipt layouts specific to regional point-of-sale systems, or how taxes and final total rounding are handled in Malaysia. No steps were taken to reduce the impact of the potential consequences listed above, but they are expected to be minor.

Additional Information

Dataset Curators

The original dataset was published by Zheng Huang, Kai Chen, Jianhua He, Xiang Bai, imosthenis Karatzas, Shjian Lu, and C.V. Jawahar for the 15th International Conference on Document Analysis and Recognition (ICDAR2019).

Licensing Information

The original dataset is licensed under the CC-BY-4.0 as noted on the ICDAR's Robust Reading Competition website for this dataset: https://rrc.cvc.uab.es/?ch=13&com=downloads.

The additions and modifications to this dataset do not alter the original license; redistributed annotations and metadata remain under the CC-BY-4.0.

Citation Information

@inproceedings{Huang_2019,
   author     = {Huang, Zheng and Chen, Kai and He, Jianhua and Bai, Xiang and Karatzas, Dimosthenis and Lu, Shijian and Jawahar, C. V.},
   title      = {ICDAR2019 Competition on Scanned Receipt OCR and Information Extraction},
   booktitle = {2019 International Conference on Document Analysis and Recognition (ICDAR)},
   publisher = {IEEE},
   pages     = {1516–-1520},
   year      = {2019},
   doi       = {10.1109/icdar.2019.00244}
}

The above reference is available at the following locations:

• arXiv:2103.10213
• doi:10.1109/icdar.2019.00244

Contributions

• Additions and minor fixes to this dataset have been made by @jsdnrs.

Acknowledgements

TODO: Add acknowledgements to other existing datasets.