Warped document restoration by recovering shape of the surface

Abstract

Document images produced by scanner or digital camera, usually suffer from two main distortions: geometric and photometric. Both of them deteriorate the performance of OCR systems. In this paper, we present a novel method to compensate for undesirable geometric distortions aiming to improve OCR results. Our methodology is based on low cost transformation which addresses the projection of curve line to 2-D rectangular area combined with finding text lines. Experimental results on several document images, indicate the effectiveness of the proposed method.

An Efficient Method for Document Correction Based on Checkerboard Calibration Pattern

“…The proposed results are compared to the state-of-the-art methods in [13,36] using OCR performance as a quantitative analysis. The most popular method for evaluation is to use OCR performance (error rate Er).…”

“…As shown in [12], these techniques rely on 2D-form recovery employing a curved surface-projection map to restore a rectangular 2D region. The uniform parameterization procedure in [13] is guided by a physical pattern. The authors of [14] use a technique that separates the page into three vertical sections and projects each one separately.…”

Warped Document Image Correction Based on Checkboard Pattern and Geometric Transformation

“…9.d shows the resulted dewarped document images using method in [14]. The results of the recent method in [14] solved the problem of the document that contains figures but removed the title of the figure. Furthermore, it failed to dewarp the equation and text in the left image.…”

“…Figure 9. The comparison between our approach and other state of arts methods using CBDAR dataset as in [14]. We compare our proposed method with the available comparison between state of arts methods [7][44] [45][4] as in [4] using ABBYY Fine Reader.…”

“…The physical pattern-based algorithm in [13] directs the uniform parameterization. The approach in [14] divided the text into three vertical sections and projected each part separately. The methods suggested by [15][16] have been used to detect the texts to describe the curved shape by using the neighboring model of transformation.…”