Improving the Accuracy of Ballot Scanners Using Supervised Learning

Most U.S. voters cast hand-marked paper ballots that are counted by optical scanners. Deployed ballot scanners typically utilize simplistic mark-detection methods, based on comparing the measured intensity of target areas to preset thresholds, but this technique is known to sometimes misread "marginal" marks that deviate from ballot instructions. We investigate the feasibility of improving scanner accuracy using supervised learning. We train a convolutional neural network to classify various styles of marks extracted from a large corpus of voted ballots. This approach achieves higher accuracy than a naive intensity threshold while requiring far fewer ballots to undergo manual adjudication. It is robust to imperfect feature extraction, as may be experienced in ballots that lack timing marks, and efficient enough to be performed in real time using contemporary central-count scanner hardware.