Document image super-resolution reconstruction based on clustering learning and kernel regression

Abstract

There are lots of blank areas and similar or redundant characters in document image. To make use of these characteristics, we propose a weighted kernel regression super-resolution reconstruction model based on steering kernel regression and clustering learning methods in this paper. By this model, we can learn the local structure of characters and achieve document image super-resolution reconstruction. In our method, a large number of unrelated samples are used for local structure clustering, which make the reconstruction process can not only use structure information of local neighborhood, but also make use of lots of non-local neighborhood structure information learning from the cluster sub-sample sets. This proposed approach ensures robustness of reconstruction. Document image super-resolution experiments with subjective evaluation and objective indicators have proved the effectiveness of our method.