Automatic Zip Code Reader

Computers excel at performing quantitative and repetitive tasks quickly, but they struggle with qualitative tasks, such as handwriting recognition. This is because it is effectively impossible to program recognition of millions of different handwriting styles into one algorithm. However, with the advent of machine learning, computers can be trained to recognize unknown objects and characters by analyzing a large labelled database. A practical use for machine learning is recognition of handwritten zip codes, which can be performed by a computer and image scanner much faster than a human can, in a setting such as a post office. In this project, the machine learning was performed within MATLAB using support vectors to define boundaries in multidimensional space between handwritten digits. The boundaries were based on a labelled dataset from MNIST (Mixed National Institute of Standards and Technology) containing 60,000 training images. Once the classification boundaries were formed, test images from MNIST were used to check accuracy, and once high enough those learned boundaries were used to determine the number in an image from a webcam with reasonable accuracy in a controlled environment.