A new method for invariant hand-written digit recognition using neural networks

Neural networks have been previously used to solve invariant pattern classification problems. Several solutions can be used. The first one consists of a pure neural solution. The neural network learns by itself the pre-processing required and extracts local information. This method gives good results but is very difficult to tune because it needs many parameters which are gray level of image pixels for the present application. The second one consists of extracting invariant features by using analytical methods. This solution also gives good classification and improves computing time. However, the efficiency depends on the type of the used descriptors such as contour features, cavity parameters and so on. We propose here a solution based on the second approach where two types of features are used in the step of information extraction. These features are fed to a neural network which is used for partitioning the space in regions corresponding to classes, and consequently for realising the classification step in a pattern recognition system independent of certain geometric transformations of the space.