Abstract. Recognition of handwritten digits is an important task in automated document analysis. Applications have been developed to read postal addresses, bank checks, census forms, among others. The Back Propagation (BP) algorithm, the most popular method for training neural networks, is being widely applied in handwritten digit recognition. However, it is often very hard to apply BP to large-scale real applications, because BP requires a long training time. In this context, some authors proposed efficient methods based on Self-Organizing Maps (SOM). The novelty of this paper is the application of a variation of SOM (GH-SOM) to this problem and the comparison of its results with the ones obtained by a stardard SOM.
[ contact authors ]
GH-SOM is a tree-structured neural network composed from independent SOMs that is capable of representing the hierarchical relations between the input data. The size of these SOMs as well the depth of the hierarchy is determined during its unsupervised training process.
The handwritten digits in our experiments were extracted from the MNIST database . We used 2000 digits for training and 3000 for testing. The original 28x28 grey level images were resized to 16x16.