Document image segmentation by a cascade of Pseudo-Word contextual labelings

摘要

The aim of this work is to classify the document content in handwritten (H), printed (P) and noise (N). In a first step, based on smearing, writing pseudo-lines and pseudo-words are extracted. The latters are classified in (P, H, N) using SVM with a Gaussian kernel. In a second step, the context of pseudo-words is examined along their pseudo-lines, spread the type of script and correct errors. First, the word separation is modeled by a conditional random field Then, the context is extended using a cascade of contextual propagation modules. Our system achieves a very good pseudo-word classification rate for both handwritten and printed text (97.3% and 99.5% respectively) for a total of 98.7%.