Phase Based 3D Texture Features

In this paper, we present a novel method for the voxel-wise extraction of rotation and gray-scale invariant features. These features are used for simultaneous segmentation and classification of anisotropic textured objects in 3D volume data. The proposed new class of phase based voxel-wise features achieves two major properties which can not be achieved by the previously known Haar-Integral based gray-scale features [1]: invariance towards non-linear gray-scale changes and a easy to handle data driven feature selection. In addition, the phase based features are specialized to encode 3D textures, while texture and shape information interfere in the Haar-Integral approach. Analog to the HaarIntegral features, the phase based approach uses convolution methods in the spherical harmonic domain in order to achieve a fast feature extraction. The proposed features were evaluated and compared to existing methods on a database of volumetric data sets containing cell nuclei recorded in tissue by use of a 3D laser scanning microscope.