AbstractsEngineering

In Thailand’s early history, prior to the availability of paper and printing technologies, palm leaves were used to record information written by hand. These ancient documents contain invaluable knowledge. By digitising the manuscripts, the content can be preserved and made widely available to the interested community via electronic media. However, the content is difficult to access or retrieve. In order to extract relevant information from the document images efficiently, each step of the process requires reduction of irrelevant data such as noise or interference on the images. The pre-processing techniques serve the purpose of extracting regions of interest, reducing noise from the image and degrading the irrelevant background. The image can then be directly and efficiently processed for feature selection and extraction prior to the subsequent phase of character recognition. It is therefore the main objective of this study to develop an efficient and intelligent image preprocessing system that could be used to extract components from ancient manuscripts for information extraction and retrieval purposes. The main contributions of this thesis are the provision and enhancement of the region of interest by using an intelligent approach for the pre-processing of ancient Thai manuscripts on palm leaves and a detailed examination of the preprocessing techniques for palm leaf manuscripts. As noise reduction and binarisation are involved in the first step of pre-processing to eliminate noise and background from image documents, it is necessary for this step to provide a good quality output; otherwise, the accuracy of the subsequent stages will be affected. In this work, an intelligent approach to eliminate background was proposed and carried out by a selection of appropriate binarisation techniques using SVM. As there could be multiple binarisation techniques of choice, another approach was proposed to eliminate the background in this study in order to generate an optimal binarised image. The proposal is an ensemble architecture based on the majority vote scheme utilising local neighbouring information around a pixel of interest. To extract text from that binarised image, line segmentation was then applied based on the partial projection method as this method provides good results with slant texts and connected components. To improve the quality of the partial projection method, an Adaptive Partial Projection (APP) method was proposed. This technique adjusts the size of a character strip automatically by adapting the width of the strip to separate the connected component of consecutive lines through divide and conquer, and analysing the upper vowels and lower vowels of the text line. Finally, character segmentation was proposed using a hierarchical segmentation technique based on a contour-tracing algorithm. Touching components identified from the previous step were then separated by a trace of the background skeletons, and a combined method of segmentation. The key datasets used in this study are images provided by…