A Study Of Language Recognition System For Document Images

In this thesis, a document image processing system is designed and performed. It can be applied for language identification. The system consists of image preprocessing, layout analysis, and language identification. The main contributions of this dissertation include:(1) Image preprocessing. In order to detect and correct the slopy angles of an image, a method based on Hough transform is presented. To reduce the computation of Hough transform, it is modified in the following ways: an appropriate quantitative angle step is taken to decrease number of angles; a sub-region other than the whole image is used to reduce the data to be processed; "featured pixels" are extracted to reduce the data further. To improve the effect of image rectification, the areas that the original pixels occupy are used to carry out the interpolation of after rotation-blank pixels.(2) Recursive dichotomy algorithm based on projection of horizontal/vertical orientation is a practical but inefficient algorithm for document understanding. To improve the efficiency and reduce the computational complexity, the dichotomy is replaced wjth polytomy algorithm. The polytomy algorithm can be applied to sub-region selection and character lines division.(3) A new shape-based method, layout understanding based on pyramid model, is proposed to solve the problem of complex layout segmentation. Experiments show that the method can get high veracity and has highadaptability.(4) In language identification, a developed Upward-Concavity algorithm is proposed to avoid parameter measuring of character lines and deviation calculation of the parameters; an improved wavelet texture-based language identification method is proposed so that the identifying work can be more interactive and more flexible; a method, called run number algorithm, is also proposed to get higher veracity.