Kuzushiji-MNIST is a benchmark dataset for image classification in Machine learning and Deep learning. It consists of 70,000 28x28 pixel grayscale images of handwritten Japanese hiragana characters, partitioned into a training set of 60,000 samples and a test set of 10,000 samples. The dataset was created as a modern alternative to the classic MNIST dataset of handwritten digits, which has been widely used since the late 1990s to evaluate Neural network algorithms. Kuzushiji-MNIST retains the same image dimensions, data format, and number of classes (10) as the original MNIST, making it a direct drop-in replacement for researchers who wish to test models on a more challenging and culturally significant task.
The dataset was introduced in 2018 by a team of researchers from the University of Tokyo and the National Institute of Japanese Literature, led by Tarin Clanuwat and colleagues. It was released alongside two related datasets: Kuzushiji-49, which contains 49 classes of hiragana characters, and Kuzushiji-Kanji, which includes 3,832 classes of cursive kanji. The primary motivation was to address the problem of optical character recognition (OCR) for historical Japanese documents, which are written in a cursive script known as kuzushiji. This script was widely used from the 8th to the 19th century but was largely abandoned after educational reforms in 1900, leaving millions of historical texts unreadable to modern Japanese speakers.
Design and Construction
The images in Kuzushiji-MNIST were derived from a large corpus of scanned historical Japanese books and manuscripts. The researchers used a combination of automatic segmentation and manual annotation to extract individual character images from these documents. Each image was then resized to 28x28 pixels and converted to grayscale, with pixel values ranging from 0 (black) to 255 (white), exactly matching the format of the original MNIST dataset. The 10 classes correspond to the ten most common hiragana characters in the corpus, which are: 'o', 'ki', 'su', 'tsu', 'na', 'ha', 'ma', 'ya', 're', and 'wo'. These characters were chosen to balance class frequency and visual distinctness, ensuring that the classification task is neither trivial nor impossibly difficult.
The dataset is distributed in a binary format identical to the original MNIST, with separate files for training images, training labels, test images, and test labels. This compatibility allows researchers to use existing code and pipelines designed for MNIST without modification, simply by swapping in the new data files. The creators also provided a Python loader script and a detailed README to facilitate easy integration into common Deep learning frameworks such as PyTorch and TensorFlow.
Benchmark Performance
Since its release, Kuzushiji-MNIST has become a standard benchmark for evaluating image classification models, particularly in the context of Data Augmentation and Residual Network (ResNet) architectures. The task is notably more difficult than the original MNIST because the hiragana characters exhibit greater intra-class variation and visual similarity between different classes. For example, the characters for 'su' and 'tsu' are often confused even by human readers unfamiliar with kuzushiji. As a result, simple models that achieve near-perfect accuracy on MNIST typically perform significantly worse on Kuzushiji-MNIST, making it a more sensitive test of model capacity and generalization.
State-of-the-art results on Kuzushiji-MNIST have been achieved using convolutional neural networks (CNNs) with Batch Normalization and Dropout regularization. As of 2024, the best reported test accuracy is around 99.5%, achieved by an ensemble of residual networks with extensive data augmentation. However, the dataset remains challenging for lightweight models and for approaches that rely on handcrafted features, which often plateau at accuracies below 95%. The benchmark has also been used to study the effects of Learning Rate Scheduling and Weight Initialization strategies, as well as the robustness of models to label noise.
Applications and Impact
The primary application of Kuzushiji-MNIST is as a research tool for advancing OCR technology for historical Japanese documents. The broader Kuzushiji project, from which the dataset is derived, aims to digitize and make searchable the millions of books and manuscripts written in kuzushiji that are currently inaccessible to the general public. By providing a standardized benchmark, the dataset has enabled researchers worldwide to develop and compare algorithms for recognizing cursive characters, which is a critical step toward automated transcription of these texts.
Beyond its direct application, Kuzushiji-MNIST has been adopted as a general-purpose testbed in Machine learning education and research. Its similarity to MNIST makes it an accessible introduction to image classification, while its increased difficulty encourages exploration of more advanced techniques. The dataset has been cited in hundreds of academic papers and is included in popular machine learning libraries and course materials. It has also been used to evaluate the transferability of models trained on other character datasets, such as the Latin alphabet or Chinese characters, providing insights into cross-script generalization.
Related Datasets and Extensions
The creators of Kuzushiji-MNIST also released Kuzushiji-49, which includes all 49 hiragana characters, and Kuzushiji-Kanji, which covers 3,832 kanji characters. These datasets are larger and more complex, with Kuzushiji-Kanji containing over 140,000 images. They are intended for more advanced OCR tasks and for training models that can handle the full range of characters found in historical documents. Additionally, the Kuzushiji project has published a large annotated corpus of historical texts, known as the Kuzushiji Dataset, which includes full-page images with character-level bounding boxes. This resource has been used in competitions and collaborative research efforts to develop end-to-end transcription systems.
In the machine learning community, Kuzushiji-MNIST has inspired similar efforts to create culturally diverse benchmarks, such as the EMNIST dataset for handwritten letters and the Fashion-MNIST dataset for clothing images. These datasets share the same format and size as the original MNIST, allowing for direct comparison of model performance across different domains. The success of Kuzushiji-MNIST has highlighted the value of preserving and digitizing historical scripts, and it continues to serve as a bridge between cultural heritage preservation and modern Artificial intelligence research.