Research On Techniques Of Script Identification Of Document Images

With the rapid development of the network communication and the information processing technology, document image is becoming an increasingly important media for the communication. Document images are derived from image data acquisition equipment such as scanner and digital camera. They are digital images coded with still image coding schemes and carry abundant text information, image information and format information.As the intercommunications among countries are more frequent and the pace of globalization has been accelerated, there are many more scripts need to be identified and processed. Script identification extracts the underlying features that can be used by the computers to automatically classify the images by scripts, and it is significant for attaining information from document images effectively.This dissertation focuses on the script identification for the document image, and proposes several new effective algorithms according to the fact that the stroke direction distributions,texture directional information and structure of of different scripts are different. The contributions of this dissertation can be summarized in the following 4 aspects:1. The development and state of art of the script identification are introduced, and based on the analysis of the script identification technologies existed, the problems are pointed out. And the structure and texture features of document images are studied thoroughly.2. A script identification algorithm based on stroke direction histogram is proposed. Considering that the stroke direction distributions of the document images in different languages are different, the edge pixels of the strokes are used to describe the stroke direction. Firstly, the distributions of the edge pixels are derived, then the features indicating the stroke direction distributions of different languages are extracted. And Support Vector Machine (SVM) is used to classify different languages. The proposed method is has great stabilization, and it can perform well even when the number of the training images is small.3. A script identification algorithm based on Brushlet transform is proposed. Brushlet transform is sensitive to directional information of image, and each sub band of document images which is decomposed by Brushlet indicates the texture distributions of the corresponding direction. Considering the fact that the texture distributions of different scripts vary, the document images are decomposed by Brushlet transform. And the energy of the half of subbands after decomposition is used to identify the scripts. Experimental results show that the algorithm proposed can identify scripts accurately.4. A script identification algorithm based on Basic Image Features (BIFs) is proposed. BIFs is a multi-scale texture analysis method, and it classified the local image symmetry of the image into 7 types, with which it describes the texture of the image. As the texture composition of document images are different when they contain different scripts, based on the theory of structural texture analysis, the 7 types of local image symmetry are taken as 7 kinds of texture basic units to represent the texture composition of document images. The proportion and co-occurrence matrixes of these texture basic units are the features for the script identification. This method can identify scripts accurately and efficiently, and is robust to degraded images.Finally, the research work for this thesis is summarized and the further research topics and directions of script identification for the document image are discussed.