Optimizing text detachment from the document image using block-based segmentation and wavelet transform

Text detachment from the background is a challenging problem in document image processing. There are many techniques for this goal. Two common methods to detach text from the background are block-based segmentation using histogram of local data and AC coefficients. Unfortunately, these methods are not accurate enough to detect block type due to noise and various distribution of the intensity values. In this paper, text detachment from the background that uses histogram of local data, is optimized by means of biorthogonal wavelets. Effect of different wavelets, their orders, and block size on accuracy is analyzed. Experiments show that using 4 × 4 document image blocks and bior3.1, bior3.3, bior3.5, and bior3.7 for feature extraction, increase accuracy incredibly. The optimized algorithm is applied to some parts of images that are extracted from English and Persian documents. We have appropriated rules of block type detection according to features of the database.