Research On Retrieval Of Historical Mongolian Document Images

With the development of digital technique, more and more historical Mongolian documents are converted into digital images so as to protect them as long as possible. This dissertation devotes to research the retrieval technology for the historical Mongolian document images, which is able to provide a way to browse and search these images by Internet. In this way, it can improve the utilization efficiency of the historical Mongolian documents. Meanwhile, it is meaningful to inherit and develop the Mongolian culture. However, the historical Mongolian documents are aging due to passage of time. Thus, the scanned images are degraded because of noise. Moreover, the same words of the historical Mongolian documents have difference in handwriting. Therefore, the retrieval technology for the historical Mongolian document images is a chanllenging task according to the above-mentioned diffculties. In this dissertation, the Mongolian Kanjur is taken as the handling object, which is a representative historical Mongolian documents. In order to solve the above diffculties, the corresponding research has been done under the technical framework of the word spotting. The main contributions of this dissertation are listed as follows:(1) In order to solve the variations in handwritting, four kinds of profile features are extracted from per image row and column of a word image respectively. Thus, four row profile-feature vectors are formulated and theirs lengths are equal to the height of the word image. And four column profile-feature vectors can also be formed and theirs lengths are equal to the width of the word image. Each word image can be represented by those profile-feature vectors. By comparison experiments, the performance of each profile feature has been evaluated. The combination of the profile features for describing each word image of the Mongolian Kanjur has been determined.(2) In the large scale of document image retrieval, the demand for the real-time respondence should be ensured. In this dissertation, the discrete Fourier transform is performed on each profile-feature vector of a word image. In the frequency domain space, the profile-feature vector can be reconstructed by a certain number of lower-order complex coefficients. In this way, each profile-feature vector will be converted into a certain number of lower-order complex coefficients. As a result, a word image can be represented by a fixed-length feature vector and each component is the modulus of the corresponding complex coefficient. Thus, the image-to-image matching can be realized by calculating the Euclidean distances between the fixed-length feature vectors of the word images, which meets the needs of the real-time in the large scale of document image retrieval.(3) An approach to generate query word images has been proposed in this dissertation, which avoids selecting an instrance of the query keyword from a large collection during the retrieval. The proposed approach has determined a set of glyphs according to the characteristics of the classical Mongolian. Based on the spelling rules of the Mongolian language, all glyphs of one word need to be connected together along with its writing order to form the corresponding word image. By this way, any word image can be generated. From the experimental results, the proposed approach not only provides convenience for users at the retrieval stage, but also ensures a stable performance.(4) According to the word inflection of Mongolian language, a method for removing inflectional suffixs from word images has been presented. For every word image of the Mongolian Kanjur, if the word image contains some kind of inflectional suffix, the inflectional suffix can be removed using the proposed method. During the retrieval, as long as the query word image contains some kind of inflectional suffix, the inflectional suffix needs to be removed from the query word image. And then, the remaining part is retrieved. Thus, the relevant words in the other person or tense will be included in the retrieval results. Experimental results indicate that the proposed method can remove inflectional suffixes from word images and partly improve the related performance.