Local Invariant Feature Histograms for Texture Classiication

This paper presents a method for texture classiication based on invariant gray scale features. These features remain constant if the images are transformed according to the action of a transformation group. The basic method applied for extracting invariant features, is given by an integration over the transformation group. For the transformation group of planar or Euclidean motion (translation and rotation) one can show, that the integration can be split into two parts: The rst is the evaluation of a nonlinear local function for every pixel of the image, and the second the summing of the results of these local computations. Instead of the second step we calculate a histogram of the local computations which preserves the invariance property and is more robust to real texture deviations than a single feature. Furthermore in a multidimensional histogram approach the combination of diierent features can be performed, thus increasing the discrimination power.