Geometric Rectification Of Camera-captured Document Image

As the showing up of OCR(Optical Character Recognition), characters in imagescaptured by fixed equipments such as scanner can be recognized and saved in computer.This operation can replace human labor. Because of the high speed and precision ofOCR, the technology is widely used in the field of computer science. In the recentyears,hand-held devices such as high resolution cameras and cell phones exhibit greatadvantages against the flatbed scanners in digitalization of documents. But due to theuncertain environment, the images captured by these portable devices always suffer fromvarious distortions. This unwanted distortions will cause serious problems OCR systemand thus should be rectified at the first time.The distortions caused by hand-held devices can be divided into two kinds,geometric and photometric distortions. This article puts more emphasis on the first one.By analyzing pictures in dataset of CBDAR2007dewarping contest,we propose adocument image rectification method based on text-lines. First of all, preprocessing ofdocument images should be done,which includes binarization, linear rectification,text andnon-text region segmentation. For binarization, a local binarization method calledSaulova is applied, which can achieve a satisfactory result on uneven illumination. Forlinear rectification, linear projection is employed. And for text and non-text segmentation,a two-class classifier is trained with svm. The second, connected components(CCs) onimage with text only are firstly extracted. Then clustering of CCs is introduced tocluster CCs into text-lines.After that, we use B-spline fitting to estimate the upper andlower baseline of text-lines.The last, coordinate transformation of the text between twobaselines is employed to rectify the text.