Digit character recognition based on Wiener filter, Karhunen-Loeve transform and BP network

A method to recognize digit characters in intensity images is provided based on traditional way of template matching. After operation of Wiener filtering on a lot of sample image templates, Karhunen-Loeve transform has been used to extract features and describe the high-dimensional images with low-dimensional matrices. Then these vectors in the low-dimensional space were loaded onto the input layer of BP network and started training. Weights were adjusted until a stable status was reached, and when preprocessing intensity images to be recognized were loaded onto the input layer, recognition results were obtained at the output layer. The Wiener filter has a good performance in recovering original signal with minimum mean square error, K-L transform can reduce the dimensionality of eigenspace and BP network does well in data mapping.