Clustering low quality Farsi sub-words for word recognition

OCR of low resolution documents is not so common, because it has a lot of problems. However, today there are several archives of digital documents which are scanned at low resolution, to consume less storage. These documents which usually have a resolution of 100 to 150 dpi, require to be converted to searchable documents. In this paper presents a new method for clustering of low quality printed Persian sub-words. This is necessary to reduce the number of classes of sub-words in order to improve the overall recognition rate. Two popular clustering methods, hierarchical and k-means implemented and compared. Local binary patterns (LBP) and zoning algorithms used for feature extraction. Both features are fast and represent the global shape information very well. Moreover, we used different distance measures to find the similarity of feature vectors. We applied our algorithms on a dataset of 10,700 images of distinct Persian sub-words with 96 dpi resolution. Experimental results show that the hierarchical clustering with the correlation distance measure has the best performance over other clustering methods and distance measures.