A simplified quantization rate-distortion model for fast document image segmentation

In this paper, we develop a new block-based multiscale document image segmentation algorithm, in which the rate-distortion (R-D) characteristic is employed by introducing a simplified quantization R-D (SQRD) model. From SQRD, an efficient entropy-based document image analysis approach is proposed. For different classes in a document image, i.e., background, text and picture, Gaussian priors are assumed and experimentally specified to characterize entropy statistics. Hereby we incorporate the entropy-based statistical document characterizations into the framework of multiscale Bayesian segmentation. The experimental results show that the entropy-based document analysis approach is promising to provide efficient and accurate compression-oriented segmentation results.