Convolution: Reading US License Plates from Multiple Perspectives

Abstract

Images with characters and numbers can be too far away for the human eye to read. This research attempts to take blurry photos of license plates and give the plate number and the associated state. Last semester we were in a car driving from Kalispell back to Helena and there was a white truck swerving on the road. Because of how fast we were traveling, none of the people in the car could read the plate number to report the truck. Some of us tried to take a zoomed-in picture to try to make the plate easier to read. This made us wish there was an application that could tell us the plate number and the state. This led to training a neural network on multiple images of state license plates. These images are randomly warped and blurred to imitate multiple perspectives. This network would be a convolutional neural network that took the large pictures of the license plate and condensed them into smaller images that are the essence of a license plate. Using this condensed data the neural network can predict the plate number and the state the plate is from. To measure how accurate this model is the images will be split into a training set and a testing set the testing set will never be seen by the model. The model’s accuracy will be determined by how well it predicts the unseen data.