Resim tabanlı osmanlıca belgelerde sınıflandırma

Abstract

Aim of this work is developing a model which classifies image-formatted Ottoman records by their contents. For this purpose, an effective classification method, which conjunctively uses "Image Processing", "Clustering" and "Natural Language Processing" techniques, is proposed for image-formatted scans of Ottoman printed records. In our work, Ottoman record samples from the official web page of Turkish Grand National Assembly (TBMM) Library and Documentation Center were used as data. Records were converted into digital form via image processing techniques, then words or letter groups in documents were detected and stored separately as individual pictures. By clustering between these pictures, identical (or similar) letter groups were registered to the same cluster. By using cluster information of letter groups, text-formatted counterparts, which include consecutive label numbers, were obtained for records. After that step, word frequency analysis, which is a valid technique in document classification, was used on converted text files as cluster frequency analysis. As a result, image-formatted scans of Ottoman records were classified based on similarity criteria of constituting letter groups, without using semantic analysis. Project was developed on MATLAB environment and classification results were obtained by a machine learning application software, WEKA. Another classification method based on word frequency analysis was also implemented using the same data set.