Number Recognizer

Documentation Status

A software for the recognition of handwritten numbers.

Requirements

  • OpenCV
  • TensorFlow
  • Numpy
  • Pandas
  • Click

Recognition Process

_images/docs.png
  • Get a grayscale version of the frame;
  • Select a ROI;
  • Apply a Gaussian Adaptive Threshold;
  • Apply Median Blur to reduce noise;
  • Apply Dilation also to reduce noise;
  • Apply an AND operation on 3 consecutive frames to reduce noise;
  • Find ROI Contours;
  • Generate a square image that is white inside the contour and black outside, for each contour found.
  • Resize image using linear interpolation to a 28x28 image, for each contour found;
  • Run an Deep Convolutional Neural Network trained for MNIST dataset on each image. For more information on the model check here.
  • Concatenate de digits and print the number.

Credits

Tools used in rendering this package: