Research on Translation and Rotation in Handwritten Digits Recognition

Handwritten digits recognition is a classical problem in machine learning and neural networks. It plays an important role in pattern recognition. In the traditional BP Neural Network, the recognition accuracy rate will drop significantly while translating or rotating test images. However, when extracting the features of the original images as the input in which the neural network is most interested, the performance of the network is improved. This paper proposes two new methods of processing the original images and extracting the feature. One is to add a sliding window and translate appropriate units to center images. The other is to rotate the input image to align its axes to its principle directions. This is achieved by Principal Component Analysis(PCA) to the input image and find its principle directions. According to the experiment results, the problem as above has been solved, and the recognition accuracy rate has improved from 96.66% to 97.12%. It is proved that the methods are simple but effective.