Rotation-Invariant Texture Segmentation using Continuous Wavelets

A successful class of texture analysis methods is based on multiresolution decompositions. Especially Gabor lters have extensively been used [1] [2] [3] [4] [5] [6]. More recently, decompositions with pyramidal and tree structured wavelet transforms have been proposed [7] [8] [9] [10]. An important aspect is the rotation invariance of the features. A discrete wavelet transform does not provide a su cient angular selectivity (only horizontal, vertical and diagonal directions). Several approaches have been studied which obtain rotation invariant features by means of interpolation [11] or data resampling [12]. Building rotation invariant features directly would lead to a robust segmentation scheme. For this purpose, we use the continuous wavelet transform (CWT), which is better suited than discrete wavelets. The most important reason for this is the ability to distribute directional information in a continuous way. By integrating over all directions, features can be obtained which are less dependent on direction than in the discrete case. An additional advantage is that less constraints are imposed on the transform. We present two kinds of rotationinvariant texture feature extraction. The rst one is based on isotropic wavelets and the second one on anisotropic wavelets [13]. The strategies are applied to unsupervised segmentation.