Translation, Rotation, and Scale Invariant Character Recognition using Modified Ring Projection

1802 Words 8 Pages

Optical Character Recognition (OCR) has been a topic of interest and research for well over a half-century, but Invariant Character Recognition (ICR), which is the recognition of characters written in different positions, orientations, and scales is still a challengeable problem. So far, many research groups have proposed different ICR techniques in the literatures. These methods can be generally divided into five groups: ‎[1],‎[2] Optical techniques, ‎[3],‎[4],‎[5] boundary-based analysis especially via Fourier descriptors, ‎[6],‎[7],‎[8],‎[9] neural network models, ‎[10],‎[11],‎[12],‎[13],‎[14],‎[15],‎[16] invariant moments, ‎[17],‎[18],‎[19],‎[20] and finally, genetic algorithms. ‎[21],‎[22],‎[23],‎[24]
…show more content…
Such networks can be fed the data from the processing of the input images and trained to output characters in one or another form. ‎[13] In particular, some neural networks use a set of desired outputs to compare with the resulted outputs and then, compute an error (deviation) to adjust their weights. In spite of the many advantages of the neural networks (such as flexibility) over conventional techniques, the main disadvantage is that they are unable to cope with large translations and rotations in the images. Furthermore, the volume of computations usually increases almost exponentially as the number of characters to be recognized is increased, making the analysis process very slow. ICR methods based on the moment invariants and invariant functions of the moments are the other popular technique used to recognize characters. Moment invariants are features based on the statistical moments of characters’ shape, which are widely used for invariant character recognition. Classical moment invariants were introduced by Hu ‎[17] and successfully used in numerous applications not only for character recognition, but also for classification and restoration of characters; nonetheless, the low-order moments contain less information about the character details and high-order ones are vulnerable to noise. The last method proposed to deal with the problem of ICR is genetic algorithm, which is a stochastic search technique used for feature selection. In the genetic

Related Documents