The module had an overall accuracy of 84.62%, with R-values for identified letters averaging to 0.6282 as shown in the table below. The left figure below shows an example of a test character and its pre-processed image as well as the template character. Furthermore, we created a dataset with a heavy amount of noise within the image and processed it through our module. This module had an accuracy of 76.92% and an average correlation of 0.6753, seen on the right of the table. The results show that the module can perform well with heavily noised images as well, and produce results similar to the original low noise samples.
(left) low noise sample; (right) salt and pepper noise sample.
Left column is the image letter, middle column is the identified character, right column is the correlation value.
Figure 4a shows the original, low noise, image before and after processing as well as the letter identified by the module. Figure 4b shows an example of the noisy image, and its counterparts after pre-processing and its identified letter. In addition, it is important to note that our original test set was written on lined paper, and our filters were successful in eliminating that along with any noise present.
(Left) original image; (Middle) image after merging & binarization; (Right) identified letter
(Left) original image with noise; (Middle) image after merging, binarization, and denoising; (Right) identified letter