Projection, smearing and Hough-based methods, classically adapted to straight lines and
easier to implement, had to be completed and enriched by local considerations (piecewise
projections, clustering in Hough space, use of a moving window, ascender and descender
skipping), so as to solve some problems including: line proximity, overlapping or even
touching strokes, fluctuating close lines, shape fragmentation occurrences. The stochastic
method (achieved by the Viterbi decision algorithm) is conceptually more robust, but its
implementation requires great care, particularly the initialization phase. As a matter of fact,
text-line images are initially divided into mxn grids (each cell being a node), where the values
of the critical parameters m and n are to be determined according to the estimated average
stroke width in the images. Representing a text line by one ormore baselines (RA method,
minima point grouping) must be completed by labeling those pixels not connected to, or
between the extracted baselines.The recurrent natureof the repulsive-attractive method may
induce cascading detecting errors following a unique false or bad line extraction.