Signature Segmentation from Machine Printed Documents Using Conditional Random Field

摘要: Automatic separation of signatures from a document page involves difficult challenges due to the free-flow nature handwriting, overlapping/touching signature parts with printed text, noise, etc. In this paper, we have proposed novel approach for segmentation machine signed documents. The algorithm first locates block in using word level feature extraction. Next, strokes that touch or overlap texts are separated. A stroke classification is then performed skeleton analysis separate overlapping text signature. Gradient based features and Support Vector Machine (SVM) used our scheme. Finally, Conditional Random Field (CRF) model energy minimization concept on approximated labeling by graph cut applied label as "signature" "printed text" accurate signatures. Signature experiment "tobacco" dataset1 obtained encouraging results.