Recognition of Laser-Printed Characters Based on Creation of New Laser-Printed Characters Datasets

Abstract

We report a new method for the recognition of laser-printed characters. We create new datasets to confirm the accuracy of our methods. When creating our datasets, printing counts are controlled, and printed character type is standardized. We purchase eight brand new manufactures laser printers and printed characters on one thousand paper sheets at each manufacture brand. We chose pages 1, 2, 299, 300, 499, 500, 799, 800, 999, and 1000 from one thousand printed sheets. One hundred printed characters on each paper sheet were captured with a high-resolution CCD camera attached with a microscope and are arranged to datasets. We extracted contours of the characters’ images of the datasets and traced x and y coordinates on the contours. Each of the coordinate was considered as a wave and was analyzed with eighth scale wavelet decomposition. We recognize between the eight brands of laser printers based on the decomposition result. In the learning phase, two hundred numbers of ‘a’ and forty four numbers ‘e’ characters printed on each sheet was leaned while in the testing phase remains of eight hundred numbers of ‘a’ and forty hundred ‘e’ characters printed on the same sheets are tested. The result shows the error rate (ER) of recognition is 2.78\(\%\) using the subspace-based method. In addition, the ER of recognition is 5.67\(\%\) using a support vector machine and 2.56\(\%\) using a convolutional neural network. As a result, we confirm that our method is able to recognize manufacturers and brands of the laser-printed characters using our datasets.