Kazakh Banknote Image Dataset
Description
This dataset contains images of Kazakh banknotes across seven denominations: 500, 1000, 2000, 5000, 10000, 20000, and a mixed category (Mix) containing multiple denominations within the same image. Each denomination folder includes at least 100 images, with the exception of the Mix folder, which contains more than 1000 images. The banknotes are photographed against various backgrounds, providing diversity in lighting and scene conditions. The 5000 KZT denomination is further divided into three subcategories: - 5000_new_single — each image contains a single new banknote, - 5000_new — images with multiple new-design banknotes, - 5000_old — images with multiple old-design banknotes. Most images contain multiple banknotes of the same denomination, except in the 5000_new_single subset. File names follow a numeric sequence (000001.jpg, 000002.jpg, …). The dataset is designed for research in computer vision, including: - banknote classification, - object detection in images with multiple notes, - studies on dataset diversity, transfer learning, and robustness of deep learning models in currency recognition tasks. All images were collected and curated specifically for research purposes. The dataset does not contain any personal or sensitive information.
Files
Steps to reproduce
1. Download the dataset archive from this repository (or clone the dataset folder structure). 2. Unpack the archive — the folder structure will be: splits/ images/ ├─ 500/ ├─ 1000/ ├─ 2000/ ├─ 5000/ │ ├─ 5000_new_single/ │ ├─ 5000_new/ │ └─ 5000_old/ ├─ 10000/ ├─ 20000/ └─ Mix/ 3. File naming: All images are provided in .jpg format with sequential names (000001.jpg, 000002.jpg, …). 4. Manifest and splits: use the provided manifest.csv for file-level metadata and the splits/ folder for predefined train/validation/test partitions. 5. Load the dataset in your preferred deep learning framework (e.g., PyTorch, TensorFlow, Keras, OpenCV). - For classification: the folder name corresponds to the label (e.g., images/1000/ → class = 1000 KZT). - For detection: use Mix/ and multi-note images as input; bounding box annotations are not provided, but can be generated with semi-automatic labeling tools. 6. Train / evaluate: models can be trained for either multi-class classification or detection tasks using the provided splits or custom partitions.
Institutions
- Suleyman Demirel atindagi Universitet