deletions | additions
diff --git a/related.tex b/related.tex
index 99da496..5f5fc1d 100644
--- a/related.tex
+++ b/related.tex
...
\section{Literature review}
\label{sec:lit_rev}
Optical character recognition has long been of interest in the computer industry. In 1976, Ray Kurzweil presented
in \cite{schantz1982history} an omni-font OCR engine, that together with a speech synthesiser would read texts to blind people. It came together with its own special scanner, that was required for taking the images of pages.
\cite{schantz1982history}
R. Holley present in a study\cite{Holley_2009} done to evaluate OCR on australian newspapers that the accuracy of various commercial software on newspapers varies from 71\% to 98\%. A similar study done by E. Klijn on dutch newspapers reported accuracies varying between 68\% to 99\%, with this number depending mostly on the quality of the image scan and on the state of the newspaper.
...
\begin{tabular}{ll}
\hline
Classifier & Test error rate \\ \hline
1-layer neural network & 12.0
\% \\
K-nearest-neighbors & 5.0
\% \\
PCA + quadratic classifier & 3.3
\% \\
SVM with Gaussian Kernel & 1.4
\% \\
2-layer neural network & 4.7
\% \\
3-layer neural network & 2.45
\% \\
Convolutional neural network & 0.95
\% \\ \hline
\end{tabular}
\end{table}
Many improvements have been published on the MNIST dataset. Some of the more recent ones include L. Deng and D. Yu\cite{deng2011deep}, obtaining an error rate of
0.83, 0.83\%, using a deep convex neural net with unsupervised pre-training. The best result obtained without using committees is D. Ciresan's deep neural network, trained on GPU devices\cite{Cire_an_2010}, that obtained 0.35\% error rate. The current state of the art was obtained by the same group, using a committee of 35 convolutional neural networks, with an error rate of 0.23\%, which beats human performance\cite{2012arXiv1202.2745C}.
For the character segmentation problem, one approach proposed by S. Lee and D. Lee\cite{Dong_June_Lee} was based on vertical projections of the images. The projections were obtained by counting how many black pixels were in a column and then various criteria were used to determine which columns were candidates for segmentation, based on some threshold values. They obtained results ranging from 85\% to 98\% accuracy, depending on font and alphabet that was used.
F. Kahraman and B. Kurt used a nonlinear vector quantization to perform the character segmentation\cite{kahraman2003license} with an accuracy of 94.5\% on a set of 2198 license plate characters.
F. Vojtěch and H. Václav presented
in \cite{Franc_2005} an approach using Hidden Markov Chains to model the problem of character segmentation in license plates. Without incorporating prior knowledge about the license plate, they get 37\% incorrect segmentations, but by using that knowledge, they managed to get it down to
3.3\%\cite{Franc_2005}. 3.3\%.
B. Janssen and E. Saund have worked on a system called Receipts2Go\cite{janssen2012receipts2go} for extracting information from small documents, including receipts. They presented two improvements to the image normalization process, but for the actual OCR part they used off-the-shelf commercial software.
