Multi-stage iterative thresholding for binarization of documents

CREATED: 201003241208 Speaker: Ruo Han

Features: pixel intensity, SNR, stroke connectivity

** Global method Otsu’s method gives one single threshold for entire image that maximizes cross-class variance

** Adaptive methods Niblack’s method: Thresold(x,y) = Mean(x,y) - k * Std(x,y) (recommended k to be 0.2) Sauvola et al: T(x,y) = Mean(x,y) (1 - k (1-Std(x,y)/R)) Gatos et al: iteratively use Sauvola’s method to estimate background and foreground

• empirically defined local window size
• need to set k
• global k for all windows

** Interactive method

• global estimate of background using Otsu’s method
• T(x,y) = M(x,y) - global background std
• User markup of bad region, expanded using MRF, re-apply the above method within that region
• Could bad region be detected automatically?

** Multi-stage iterative method

• more background pixels than foregound, use of background estimate is good approximation
• wide spread of foreground intensity, adaptive method is necessary
• varying foreground intensity even within small region, foreground pixels may be shadowed by its peers