EMNIST (Extended MNIST) is a dataset of handwritten characters that extends the original MNIST dataset of digits to include both digits and uppercase and lowercase letters. It contains 240,000 training images and 40,000 test images, each a 28x28 grayscale image. EMNIST is widely used in Machine learning and Deep learning research for tasks such as character recognition and benchmarking classification algorithms.
The dataset was introduced in 2017 by Gregory Cohen, Saeed Afshar, Jonathan Tapson, and André van Schaik at the MARCS Institute for Brain, Behaviour and Development, Western Sydney University. It was created by converting the NIST Special Database 19, which contains handwritten samples from over 3,600 writers, into the same format as MNIST. EMNIST provides a more challenging benchmark than MNIST because it includes letters, which have more classes and greater visual similarity between certain characters (e.g., 'O' and '0', 'I' and 'l').
Variants and Splits
EMNIST is available in several variants, each with a different class structure. The most common is EMNIST Letters, which contains 26 classes (A-Z) and 145,600 training images. Other variants include EMNIST Digits (10 classes, 280,000 training images), EMNIST Balanced (47 classes, 112,800 training images), EMNIST ByClass (62 classes, 697,932 training images), and EMNIST ByMerge (47 classes, 697,932 training images). The Balanced and ByMerge variants merge certain letter cases (e.g., 'C' and 'c') to reduce ambiguity, while ByClass keeps all classes separate.
Usage in Research
EMNIST has become a standard benchmark for evaluating Neural network architectures and training techniques. It is often used to test improvements in Data Augmentation, Dropout, Batch Normalization, and other regularization methods. Researchers also use EMNIST to study transfer learning, where models pretrained on larger datasets are fine-tuned on EMNIST. The dataset is included in popular machine learning libraries such as PyTorch and TensorFlow, making it easily accessible for experimentation.
Relationship to MNIST
EMNIST is designed as a drop-in replacement for MNIST, with the same image size and format. However, the inclusion of letters increases the difficulty: the number of classes rises from 10 to up to 62, and the inter-class similarity is higher. This makes EMNIST a more realistic benchmark for handwriting recognition in real-world applications, such as postal mail sorting and form processing. Many studies report that models achieving near-perfect accuracy on MNIST see a significant drop on EMNIST, highlighting the need for more robust methods.
Limitations and Extensions
Despite its wider scope, EMNIST still has limitations. It is derived from a single source (NIST), so it may not capture the full diversity of handwriting styles across different populations. The images are also preprocessed to be centered and normalized, which reduces variability. To address these issues, researchers have created extensions such as Fashion-MNIST (clothing images) and Kuzushiji-MNIST (Japanese characters), but EMNIST remains a widely used baseline. As of the early 2020s, EMNIST continues to be cited in hundreds of papers annually, and it is often used in educational settings to teach Deep learning concepts.