Rotation Invariant Pattern Recognition with Multiscale Smoothing, Ridgelet-Fourier Features and Voting Technique

In this paper, we propose a novel method for rotation invariant pattern recognition. We perform smoothing with different filter sizes, extract ridgelet-Fourier features from each smoothed image, classify the pattern with each of the extracted feature vectors, and conduct majority voting to determine the final class label of the unknown pattern image. Our new method is invariant to the rotation of the pattern images, and it has the capability to reduce a significant percentage of random noise from pattern images. Experiments demonstrate that our new method outperforms the ridgelet-Fourier descriptor and the Fourier-wavelet descriptor for both a printed Chinese character dataset and an aircraft dataset.