Deskewing Using Binary and Grayscale Images

Document Image Analysis has today become an increasingly important domain due to the de-

sire to reduce the amount of paper documents and archives. Optical Character Recognition

(OCR) systems and document structure analyzers are the essential tools to achieve this task.

It often appears that the document to be recognized is not correctly placed on the flat-bed

scanner, especially when the document comes from a book or a magazine. This results in a

skewed digitalized image which is a real problem for the document analysis, understanding,

character segmentation and recognition. Deskewing the input image is then a crucial step in the

document understanding. In this report, we propose a deskewing method based on the Hough

transform and filtering algorithms.

Moreover, the study of characters features is necessary if a most faithful reconstitution of the

document is expected. This implies the study of characters color, average boldness and skele-

ton. Furthermore, most Optical Character Recognition systems are sensitive to the quality and

size of the given characters. Skew angle correction and binarization steps damage the charac-

ters, especially for little characters. In order to maximize the efficiency of characters recognition,

these are upscaled and smoothed by using pixel art scaling algorithm.

This report is composed of 2 sections. Firstly, the different proposed algorithms to estimate a

document skew angle and secondly the characters features study.

In this section, we propose a simple method to estimate the skew angle of a document based

on the combination of the Sobel edge detection filter, a filtering algorithm and the Hough trans-

form. This method has been designed for a quick detection of li...

... middle of paper ...

...too many pixels are deleted during the

filtering step, which then won’t leave enough information for the Hough transform.

The tolerance of this filtering algorithm depends on the window size and the threshold value.

The more the difference between the window size and the threshold value is important, the

more tolerant the algorithm is.

1.2.2 Grayscale images filtering algorithm

This grayscale filtering algorithm has been designed to do the same type of filtering than the

previous algorithm, but on grayscale images. The major difficulty, compared to binary images,

lies in the distinction between the objects and the background. In binary images we only deal

with true or false values, while in grayscale images the values range goes from 0 to 255.

For the remaining of this paragraph we assume that the text has a darker color than the back-

ground.