Word Recognition for the Balinese Palm Leaf Manuscripts

Most of the ancient Balinese manuscript collections on palm leaves are now in danger from the natural physical degradation caused by the aging period of the leaves as the writing media and the unprotected storage from dust, fungus, and insects. The digitization projects for Balinese palm leaf manuscripts, which already collected many digitized manuscripts, were halted without the next step to create a complete transliterated text of the manuscripts. To overcome this problem, an automatic word recognition system is urgently needed to be integrated into the manuscript digitization project. This paper presents an experimental study on word recognition for the Balinese manuscripts on palm leaves. In this research, some image-based feature extraction methods are used and are tested to recognize the isolated word segment images without doing any “character-by-character” recognition step and without doing any OCR-transliteration method. The recognition rate for three different features, with six different schemes and parameters for the classifiers are calculated. The proposed feature combination of HoG with the NPW-Kirsch features improve the recognition rate, and the K-NN classifier performs better compare to the NN classifier in the condition of the imbalance number of data training for each class in the dataset. The effect of the length of the word on the recognition rate are also analyzed. We found that longer words have better recognition accuracy, both in terms of Recall and Precision values. The results of this research serve as the initial analysis for the future word spotting system for the Balinese palm leaf manuscripts.