Research On Rectification Of Visual Document Images

As we all know, traditional scanner-oriented optical character recognition(OCR) techniques have made great advances in recent years and they are also successfully transformed into useful products in consumer market. However, camera-based OCR technology is still in the experimental stage and far from the practical level. Due to the differences between cameras and scanners, traditional scanner-oriented OCR techniques are not generally applicable to camera-captured documents. Unlike a large-sized image obtained by a flat-bed scanner, a camera-captured image may suffer from problems such as perspective distortion, uneven lighting and non-planar page shape. In these cases,the texts which happens perspective distortion and twist deformation become difficultto split, resulting in failure of OCR algorithm. This dissertation summarizes the existing the existing document image restoration method, puts forward a stable and fast restoration method for the camera-based document image, and in-depth research on related technologies.Camera-based document image restoration method proposed in this dissertation involves mainly image preprocessing, text flow detecting and automatically generating of restoration network. In the section of image pre-processing, we analyze the source of the interference effect restoration, and propose a text extraction method based on connected domains according to the characteristics of interference. Experiments show that, the image pre-processing method can effectively extract text information from background interference, which reduces the complexity of text flow detection and also improves the stability of the algorithm.In the section of text flow detection, we adopt different approaches to calculate vertical and horizontal text flow. Since most of the documents not in vertical bucket deformation, we calculate the vertical text flow by detecting the text baseline on both sides, and the experiment proves that this approach is fast and stable. Methods used for horizontal text flow estimation is evenly dividing the image into multiple meshes and calculating the skew angle within each grid, and then we obtain a horizontal text flow matrix. In this dissertation, we introduce the idea of multi-resolution which usingdifferent size of meshes to calculate the text flow. The experimental results of singleresolution and multi-resolution are compared to prove the multi-resolution method can significantly improve the accuracy of detection.Based on the horizontal and vertical text flow matrix, we present an approach to generate documents outline network automatically and describe the approach of generating lateral network curve. In order to evaluate the quality of network, we propose using network covering text area ratio and lateral line deviation rate as the evaluation standard and test by software rendered image and camera-captured image. At last, we record the program running time. Experiments show that the restoration algorithm for camera-captured documents proposed in this dissertation works effectively and quickly.