Reconnaissance d'écriture manuscrite par des techniques markoviennes : une approche bidimensionnelle et générique

We present an approach of the problem of handwriting recognition using hidden Markov random fields and based on a trully bidimensional analysis of the handwriting. The main innovative aspect of this approah is the combination of a windowed analysis of the image, a Markovian modelisation and an implementation of the 2D dynamic programming algorithm that achieves a fast and optimal decoding of Markov fields. Another feature of this study is the development methodology that focuses on a systematic evaluation of the algorithms and parameters. These algorithms are partly taken from techniques of the domain of speech processing and are very generic. This approach is validated on two different applications corresponding to standard public databases. The application of this generic algorithm to a handwritten digits recognition task achieved results similar to the ones of state-of-the-art methods. The application to a handwritten words recognition task showed that this approach can be extended to more complex tasks in a natural way. This work showed that the proposed approach is valid and appears as a candidate standard method for solving various tasks in computer vision. It paves the way for further developments in handwriting recognition and other important enhancements are expected with the use of other principles commonly used in speech and language processing problems. Other tasks such as image segmentation could also benefit from the robustness and the learning ability of our approach.