Character Extraction from Interfering Background - Analysis of Double-Sided Handwritten Archival Documents

The sipping of ink through the pages of certain double-sided handwritten documents after long periods of storage poses a serious problem to human readers or OCR systems. This paper addresses this problem through the recovery of content on the front side of a page from the interfering image caused by the handwriting on the reverse side. First, by adapting the Gaussian stochastic model, the interfering model based on norm-orientation-discontinuity is proposed in analyzing the properties of the interfering strokes. Secondly, an improved canny edge detector with edge norm-orientation similarity constraint is proposed. At the same time, two low thresholds are used to detect edges instead of a single low threshold. This improvement could link weaker foreground edges without introducing noises in the overlapping/overshadowed area. The proposed algorithms perform well regardless of the intensity differences between the image on the front side and the interfering image from the reverse side. The segmentation results of real images are shown and evaluated