Use of eigenvector projections as inputs to a neural net for handwritten digit recognition
The usual approach in handwritten digit recognition based on neural networks makes use of the raw image as training and processing inputs. In this paper, instead of using the image, we obtain the principal components (PC's) or eigenvectors of an ensemble of digitized handwritten numbers from 0 to 9 and the projections of the numbers are used as the inputs to the neural net. Each digit is 6 x 8 pixels. The main advantage of using projections instead of raw images is the reduction of the number of inputs which in turn leads to reduced computational complexity. To our knowledge, nobody has used number projections as neural network training inputs yet.