Automatic segmentation and recognition of zip codes on handwritten real mail

The authors combine statistical and structural techniques for automatic segmentation and recognition of zip codes. This kind of combination has been advocated by several authors. The aim is to incorporate advantages of both approaches, that is to say, the flexibility and the adaptability to noisy patterns, for statistical techniques, and possibility to describe very complex and specific patterns, using syntactic techniques. The statistical stage is the first and main stage of the authors recognition system. It uses concavity measurements to characterize the characters and a linear discrimination technique is implemented to affect one character to one cluster. The second stage, called structural supervisor, is developed in order to control and improve the results of the first stage.<
>