This system is under development by Samia Maddouri (ENIT, LSTS, Tunisia). It is about a system for recognition of handwritten Arabic words by a transparent network of neurons. This system functions by propagations - retropagations repeated until the final choice of only one candidate to the recognition. During the movement of propagation, primitives are extracted from the word voter for the primitives which it contains. Then, the stimulated letters vote in their turn for the words which contain them. During the movement of retrpropagation, the stimulated words indicate the missing letters and provide indications on their site. This information is transmitted through the letters and the primitives stages which supplement them by indications on the size and morphology. When returned on the word, windows are placed on the letters, then the contour of each letter is normalized by approximating the corresponding printed letter, by using the Fourier descriptors.
The first stage is reserved to the characterization of Arabic script
by its contour. A mathematical description of the contour by the
Fourier coefficients is formulated and the harmonic and elliptic characteristic
of the latter is graphically illustrated. The characteristics retained
for the choice of the close characters: Hampe, Descender, Buckle and diacritic
Points. A characterization of the Arabic characters printed by the
swing angle of their first harmonic is then carried out in order to prepare
at the stage of normalization.
Character Normalization
The recognition of the writing by Fourier Transform is operated beforehand on the character contour. The following figure presents the stages of the normalization system . In the first stage, we start with the detection of contour. In the second stage, the contour Freeman code is generated, on which we operate the calculation of the Discrete Fourier Transform of Fourier Discrète (an analysis in series of Fourier is also possible).
The contour approximation by the Fourier coefficients can be seen
like a superposition of elliptic forms of various sizes, positions and
orientations. The following figure presents a graphic illustration
of ßt, 2nd, 3rd and 4th harmonic approximation of Fourier of the
contour of the character " Ba ".
Rotation angle normalization
This transformation allows to bring back the inclination angle of the
first harmonic of all the characters to 0. This is effective to eliminate
the variability of slope inter characters, but risk to reduce the distance
will intra characters. In order to avoid this risk, an idea is introduced
on this level. This idea consists in
normalizing the swing angle of the handwritten character compared to
the inclination angle of the principal axis of the first harmonic corresponding
to the reference characters . Figure (b) presents the inclination
rotation angle of the character " Ra " of the figure (a) in order to bring
back it to 0. Figure (c) presents the rotation of the same character
to align it with the principal axis of the first harmonic of the subset
of reference to which belongs the character " Ra ". I.e. to bring
back the angle to Pi/4.
Extraction of the central band
Characteristic extraction
Reconnaissance of the chosen word
