Non-Uniform Slant Correction for Handwritten Text Line Recognition

Towards robust writer verification by correcting unnatural slant

“…An alternative option is to shear text lines or words individually, but this is less reliable and breaks ink traces at region boundaries. It is also possible to eliminate slant non-uniformly within each text line, but this seems to add little or no improvement, despite the added complexity (Bertolami et al, 2007). The page-level approach is simple, fast and keeps the signal structurally intact.…”

“…It is a standard step in handwriting recognition systems, which use it to optimize recognition of the textual contents (Bertolami et al, 2007;Kavallieratou et al, 2001). It has also been called ''deslanting'', ''slant removal'', and ''slant correction''.…”

Markov models for offline handwriting recognition: a survey

“…Depending on the actual configuration of the particular experiments and the recognizer used, word error rates between 37.3% [9] and approximately 53% [117] have been achieved. Furthermore, in [117] results for evaluations on the Cambridge database (≈9% WER) were reported.…”

“…2 Schematic overview of a typical architecture for an HMM-based handwriting recognition system. Example whiteboard data, preprocessing, and recognition results shown are taken from [125] successful methods rely on gradient information extracted from the text-line image (cf., e.g., [9,38]), which is often binarized in advance.…”

Neural Networks for Document Image and Text Processing

“…This work is similar to [Slavík et al 2001] in which Sobel filters are computed to calculate the best slant angle. [Bertolami, Uchida, et al 2007] uses a local slant correction instead of averaging the slant of each word/line. An alternative is to use ML to estimate the MBA avoiding geometrical heuristics.…”