الفهرس 
Image BinarizationOnly 14 pages are availabe for public view
 |  | from | 202 |  |  |
| | | from | 202 | | |
Abstract
Efficient image binarization technique is critical for achieving high quality in the document image analysis pipeline. The binarization process is a challenging and a very difficult task because these document images may be degraded where they may be suffering from aging, leaked ink, bleeding through, stains, smudge, spots, image contrast variation, non-uniform illumination, heat damage, etc. Conventional binarization techniques do not take all these degradations into account
and therefore do not produce efficient binarization. Additionally, the global threshold techniques give unstable results and the local threshold techniques also do not give good results on degraded document images.
Furthermore, the modern binarization techniques make image binarization an interesting research subject for the optimiza- tions targeted to increase the execution performance and to improve the binarization quality.
This thesis presents an efficient multistage binarization tech- nique for the degraded document images. This binarization technique won a rank number one in H-DIBCO 2018 con- test where DIBCO contest is the most popular binarization contest in the world and it is funded by the European union. Also in DIBCO 2018 contest, this binarization technique won the rank number one for two consecutive years using one of their measured formulas for all participators in DIBCO 2018 contest and all participators in DIBCO 2017 contest.
In this technique, an edge detection method is proposed to be used for the proposed selecting seeds algorithm. Another edge detection method is proposed to improve the edges of the objects in the image. Then, to find the rest of the ob- ject’s contents, Markov random field is applied in an energy minimization framework using graph cuts. A combination algorithm is proposed between these results and Niblack’s
ix
method, where this algorithm depends on the flood fill algo- rithm. Then a proposed post processing algorithm is applied, which also depends on the flood fill algorithm to produce the binarized image.
The results of extensive experiments on many datasets show the robustness of the proposed technique on various types of degradations in the document images where the proposed technique demonstrates superior performance against many other methods.
Additionally, this thesis proves that the measurement formu- las for computing the evaluation results in the contests are inaccurate and proposes binarization measurement formulas. Also, this thesis proposes new equations for the old binariza- tion measurement formulas.
Furthermore, in the contests they use the ground truth image for the evaluation but the ground truth image itself usually has many mistakes.
This thesis presents a proposed binarization measurement al- gorithm with its proposed equations, where using this algo- rithm in the evaluation reduces the effect of the mistakes of the ground truth image in the result.