Bitmap Numbers Dataset

A dataset of labeled handwritten numbers for machine learning.

About

This dataset contains manually labelled images of hand-written digits. The root directory contains subdirectories for each labeled digit. Every image is a bitmap, and the filenames are in the following format: [image number].bmp All images are resized to 128x128. The color depth is always 24 bit.

Potential Issues

• The images were not taken at the 128x128 resolution, so scaling issues may occur.
  • Some scans did not originally have a 1:1 aspect ratio, so images may be slightly stretched.
• Some images are in color, while others are in monochrome.
  • The images were collected from different sources, so the colors may vary.
• Images may have artifacts.
  • Scans were converted from pdf to png, then to bmp. Some artifcats may appear.
• The image size may be too large.
  • 128x128 may be too large to train in a short amount of time.
    • You can easily convert these images to smaller sizes using command line tools such as ImageMagick.

Credits

All samples from this dataset are from the public domain numbers repository by @kensanata. The repository contains extra metadata such as country, gender, and age, as well as automatically categorized samples and other useful tools.