A Feature Vector for Optical Character Recognition

The extraction of the written text in an image has always been an important application of computer vision since it was introduced. It is widely used in automatic number plate recognition, handwriting recognition, extracting data from scanned documents such as passports, ID cards, banking forms, etc. There exist a wide variety of approaches to the general problem of optical character recognition such as Template Matching, Structural Classification, Artificial Neural Networks, etc. In this paper we introduced a new feature vector for optical character recognition and we tested its accuracy by using a Nearest Neighbor classifier. The new feature vector is a sequence generated by putting together the orientations of each pixel to a base point. The classifier then, is simply Longest Common Subsequence algorithm. In other words, a new image contains a character if and only if the corresponding sequence of the image has the longest common subsequence with the feature vector or sequence of that character among all the characters available. The experiments provided us with satisfying results which can be definitely better under better classifiers such as RNN or SVM.