Estimation of Tilt in Characters and Correction for better Readability by OCR Systems

Abstract

The existing Optical Character Readers (OCRs) are capable of reading linear form text and have limitations to read artistic and non-linear form text. The tilt in characters contributes a major share in affecting the efficiency of the recognition algorithms. This paper presents a technique to estimate and correct the vertical tilt in printed characters of English in order to make an OCR to read the text more efficiently. The input characters are assumed to be segmented from the document image and free from noise. Initially, the direction of tilt of the characters is detected using a heuristically constructed knowledgebase. Next, the inclination of the character to its base is estimated using line drawing algorithm. Finally, the estimated tilt is corrected through rotation in counter direction of the tilt. The method has been tested with sufficient samples and readability analysis is performed with an OCR. Experimental results show an average improvement in readability by OCR from 20% before tilt correction to 82% after the tilt correction.

A Model to Convert Wave-Form-Text to Linear-Form-Text for Better Readability by OCRS

“…Analysis of readability by an OCR of the text before transformation and after transformation is performed with respect to English text using the OCR "Readiris Pro 9" [23]. In this process, first, the samples of wave-form-text are taken as input to OCR and subjected to be read by the OCR and the result is tabulated in Table 1.…”

“…The work proposed by Vijayashree et al [23] presents a simple technique to estimate and correct the tilt present in artistic text. Initially, the direction of tilt of the characters is detected using a heuristically constructed knowledgebase.…”

Data set for Tifinagh handwriting character recognition

“…The extraction steps were: We use the horizontal histogram to correct the inclination of every page [6] . Using connected components algorithm we detect the center of each character [7] .…”

A Masked Bounding-Box Selection Based ResNet Predictor for Text Rotation Prediction

“…Research on text rotation prediction can be divided into two branches, the traditional Computer Vision methods based (Shukla et al, 2016;Pan et al, 2010;Veena M.N and Vasudev.T, 2015;Chinara et al, 2018;Vijayashree et al, 2014) and the deep neural networks based (Channa and Rao, 2020;Akhter and Rege, 2020). Shukala (Shukla et al, 2016) et al presented a Hough transform method to detect skews in licence plates.…”

“…Chinara et al (Chinara et al, 2018) proposed to use Centre of Gravity (COG) method to predict the rotation of the images. Vijayashree et al (Vijayashree et al, 2014) proposed a method to first segment the input characters and then the inclination of the character to its base is estimated using line drawing algo- Though progress has been made in the text rotation prediction task, the state-of-the-art methods are far from being satisfactory. Previous work cannot address the challenging rotation prediction problem when the background is noisy.…”