A binarization technique using local edge information

Binarization is usually the first step in document analysis and recognition. This paper propose a new binarization technique base on local thresholding using information from edge pixel. Indeed, an edge represent a rupture between foreground and background elements. It is then reasonable establishing a local threshold on each edge pixel. An average of this edge-based local threshold is used to classify each pixel as either foreground or background. This averaging process can be done rapidly thanks to the concept of integral image. The resulting algorithm is fast and contains few adjustable parameters. A preliminary experiment showed that the proposed method outperform existing method on our ancient document images collection.