alakxender/dhivehi-layout-syn-b1-florence

Synthetic Dhivehi Document Layout Analysis Dataset

Overview

This dataset contains synthetic document layouts annotated with bounding boxes and labels for various sections in the Dhivehi language. It is designed for training models on document layout analysis and Optical Character Recognition (OCR) tasks. The dataset simulates real-world documents in Dhivehi and can be used for layout-aware OCR to recognize text and understand the structure of Dhivehi documents.

Dataset Structure

Each data point in the dataset consists of:

• Image: A synthetic scanned document image (e.g., .jpg, .png) with Dhivehi text.
• Annotations: Bounding boxes and labels indicating the locations of various components of the document. The prefix and suffix in the annotations are adjusted for Florence2 style, with each label followed by its respective bounding box coordinates.

Annotation Format

Annotations are represented in the following format:

prefix<locXXXX><locYYYY><locZZZZ><locAAAA> label_name ;

Where:

• prefix: A predefined identifier indicating the document section type.
• <locXXXX>: X-coordinate of the top-left corner of the bounding box.
• <locYYYY>: Y-coordinate of the top-left corner of the bounding box.
• <locZZZZ>: X-coordinate of the bottom-right corner of the bounding box.
• <locAAAA>: Y-coordinate of the bottom-right corner of the bounding box.
• label_name: The label for the section (e.g., "heading", "caption", etc.).

Categories

The labels used for the bounding boxes in this dataset are:

• author
• caption
• columns
• date
• footnote
• heading
• paragraph
• picture
• textline

Example Annotation:

<OD><loc_131><loc_32><loc_945><loc_70> heading ;
<OD><loc_167><loc_102><loc_947><loc_141> heading ;
<OD><loc_363><loc_138><loc_948><loc_177> date ;
<OD><loc_841><loc_213><loc_952><loc_231> textline ;
<OD><loc_117><loc_258><loc_948><loc_284> textline ;
...

Synthetic Nature of the Dataset

This dataset is synthetic, meaning that the images and annotations are artificially generated to simulate real-world document layouts in the Dhivehi language. While the documents may not correspond to actual historical or published material, they follow realistic formatting and structures that reflect how text in Dhivehi would typically appear in different document sections.

The synthetic nature of this dataset allows for controlled experimentation, and it can be used to train models without the need for large collections of real-world document scans.

Dataset Details

• Language: Dhivehi
• Synthetic Data: Artificially generated document images and annotations.
• Number of Images: 1,356 images
• Image Format: .jpg, .png
• Resolution: 300dpi (example)
• Number of Annotations: 1,800 bounding box annotations (example)
• Document Type: Synthetic, Dhivehi-language documents

Dataset Splits:

• Train: 1,356 examples, 112.65 MB
• Test: 297 examples, 24.84 MB
• Validation: 298 examples, 25.02 MB

Total dataset size: 162.51 MB
Download size: 157.88 MB