Contrasting classifiers for software-based OMR responses

Systems based on optical mark recognition (OMR) provide a means to address rising student numbers in university modules. Developing a system which addresses the needs of a specific module provides added flexibility to enhance the teaching environment. The process is complicated by responses which have been selected and then deselected by the respondent.This study contrasts four classifiers for identifying selected responses on OMR answer sheets. Four classifiers are constructed based on features derived from the number of pixels in an image, patterns derived from the image, edit distances derived from bit string patterns and the average number of edges per axis. The classifiers based on the number of pixels and edges make use of simple thresholds for classification, whereas the classifiers based on patterns and edit distance make use of classification trees. The classifier based on the number of pixels in the image delivers the best results as it has very high levels of accuracy, sensitivity and specificity, but is unable to identify any responses which have been deselected by a respondent. The classifier based on the number of edges per axis performs very well in accurately classifying positive responses, but has only moderate levels of general accuracy. The edge-base classifier is, however, the only classifier able to correctly classify any deselected responses.