基于逆向工程的扭曲文档图像校正技术研究

Over the years, we have been engaged in the R & D work of visually impaired assistive products. In the R & D process, we have been unable to solve the problem of the distorted document image correction and recognition. The studies of this project is how to use the reverse engineering methods to correcting the distorted document image distortions. The main idea is to get the texture of a document image information through the texture camera, to obtain the depth information of the document image through the stereo camera and the structured light projection,and to calculate pixel spatial distance by curve integral value for the leveling treatment of distorted document image through combining image data processing method by the 2D image data and 3D image data, finally to achieve the purpose of the correcting. The key technologies includ post-processing point cloud data based on median filtering, document image tilt correction based on 3D data, the geometric model surface recovery based on the piecewise cubic spline interpolation. The goal of this study is mainly to get flat document image information through the processing distorted document images and to improve OCR optical character recognition rate. The research project will lay a solid foundation for the further application and make the the original lower distortions document recognition rate be improved.

多年来，我们一直从事视障者辅具产品化的研发工作，在研发过程中，我们一直遇到对扭曲文档图像的校正与识别的难题。本项目致力于研究如何通过逆向工程的方法对扭曲文档图像进行扭曲校正。主要思路是通过纹理相机获取文档图像的纹理信息，通过立体相机与结构光投影获取文档图像的深度信息，通过2D图像数据与3D图像数据相结合的处理方法，对扭曲文档图像的像素空间距离进行积分计算，从而对扭曲文档图像进行虚拟辗平处理，达到扭曲校正的目的。采用的关键技术包括基于中值滤波的点云数据后处理、基于3D数据的文档图像倾斜校正、基于分段三次样条插值曲面的几何模型恢复等。研究目标主要通过对扭曲文档图像的信息处理，得到较为平整的文档图像，并使得OCR光学字符识别率有所提升。本项目的研究成果将为文字识别技术的进一步应用与推广打下坚实的基础，使得原先较低的扭曲文档识别率有所提升，使得面向扭曲页面的文档图像识别技术进一步实用化。

针对数码摄像头拍摄的文档图像存在扭曲变形从而影响OCR(光学字符识别)识别率的问题，本项目研究了基于逆向工程的扭曲文档图像校正技术。整体内容可分为三部分。一、点云数据预处理部分：首先利用基于K-近邻的散乱点云去噪算法进行点云获取，随后研究了基于书籍形态特征对书籍模型进行背景去除和单页面模型提取的方法。二、图像校正处理部分：单页面模型提取后，对模型进行了倾斜校正，研究了基于三次样条插值方法和离散化近似计算方法对3D平面进行扭曲恢复的方法，并研究了基于3D模型的二维图像提取方法。三、扭曲文档图像校正的后处理部分：研究了将3D模型得到的二维图像进行自适应图像增强处理，并进行OCR识别的过程。通过以上处理，能够较好地完成基于三维模型的扭曲文档恢复，相对于2维扭曲文档的恢复，OCR识别率有所提高。本项目发表学术论文共40篇，专利4个，其中发明专利4个。所研究的算法成果已经在视障者阅读辅助器具上试用。