BD-BANKNOTES: A Multi-Series Annotated Dataset of Bangladeshi Banknotes for Object Detection

Description

The "BD-BANKNOTES" dataset has been systematically organized to advance the visual detection, localization, and multi-series classification of Bangladeshi paper currency under unconstrained real-world environments. Structured in accordance with standard YOLO framework specifications, the repository provides a unified collection of 3,466 curated high-resolution field photographs within an `images/` directory and their corresponding ground-truth annotation text files within a `labels/` directory, managed by a master `data.yaml` configuration file. Bounding box coordinates and zero-indexed denomination class labels (0 to 8 across 2, 5, 10, 20, 50, 100, 200, 500, and 1000 Taka) are preserved in normalized YOLO format, guaranteeing precise object localization and seamless compatibility with SOTA object detector architectures such as YOLO26. To support both denomination identification and cross-generational domain adaptation, the dataset uniquely encompasses older circulating banknotes (1,955 images; 56.40%) alongside the newly issued 2025-2026 "Historical & Archaeological Architecture of Bangladesh" series (1,511 images; 43.60%), which introduces authentic student protest graffiti (Graffiti-2024) and landmark national monuments. Semantic series and denomination metadata are encoded directly into each file's naming scheme through sequential indexing (`<Series>_<Denomination>_<Index>.jpg` and `.txt`), accompanied by a comprehensive `file_mapping.csv` audit trail and deterministic 80:10:10 partition manifests (`train.txt`, `val.txt`, `test.txt`). This meticulous organization facilitates reproducible benchmarking, assistive mobile vision for visually impaired individuals, and embedded cash-handling applications, making the "BD-BANKNOTES" dataset an essential, high-utility resource for researchers and practitioners in computer vision and currency analysis.

Steps to reproduce

To reproduce the BD-BANKNOTES dataset for academic and benchmarking purposes, follow this systematic workflow: Capture high-resolution field photographs of Bangladeshi paper currency across all nine circulating denominations (2, 5, 10, 20, 50, 100, 200, 500, and 1000 Taka), encompassing both the older circulating series and the newly redesigned 2025 Architecture series under diverse real-world lighting, background textures, viewing angles, and physical wear conditions; screen the raw captures to filter out severe motion blur, improper focus, and occluded frames; annotate each banknote with precise bounding boxes and zero-indexed denomination class IDs (0–8) in normalized YOLO format using AnnotateAI; systematically encode series and denomination metadata into sequential filenames (<Series>_<Denomination>_<Index>); organize the curated files into unified images/ and labels/ directories paired with a master data.yaml configuration; and partition the dataset into independent 80:10:10 training, validation, and held-out test splits using a fixed random seed (seed 42) for reproducible model benchmarking.

Institutions

• United International University

  Dhaka Division, Dhaka
• BRAC University

  Dhaka Division, Dhaka

Categories

Licence