Non-uniform skew estimation by tensor voting

We apply a perceptual grouping concept for document deskewing. The method is similar to Hough Transform, but each voter casts its votes in a neighborhood instead of the whole domain; and the winning policy is based on tensors, which are more sophisticated than scalars and vectors. The voters are centroids of connected components, and skew angles are estimated along the skewed text lines by tracing the ridges (arbitrary curves of peaks) in the voting domain, and the complete skew angle field is obtained by elastic surface interpolation.