A technique of simultaneous blind image restoration and image labeling is applied to the preprocessing of images of highly degraded ancient printed documents. The aim is to improve the visual quality of the documents themselves and to facilitate the subsequent phases of segmentation, recognition, and classification of the text characters. Following recently proposed techniques of strict Bayesian fusion of visual modules, we integrate data, a priori information, and image and degradation models in order to define a joint energy function. The minimizer of this energy is an estimate of the ideal undegraded image in an already segmented form, plus an estimate of the degradation operator. We propose a solution whereby image estimation is iteratively alternated with the estimation of the degradation operator, Several results of experiments on both simulated and real degraded images are shown to validate the method.
展开▼
