Rotated, Scaled, and Noisy 2D and 3D Texture Classification with the Bispectrum-Based Invariant Feature

摘要

The author presents a novel feature of 2D and 3D images invariant to similarity transformations and robust to noise on the basis of the bispectrum. The invariant feature is applied to the classification of texture images suffering from rotation, scaling and noise. Computer experiment shows that about 90 % correct classification ratio is obtained for 5 kinds of 2D natural textures and of 3D brain images rotated in arbitrary degree, scaled up to double and with the white Gaussian noise of 0 dB SNR. The feature can also be used to the estimation of the rotation angles of texture images.