Material Recognition for Content Based Image Retrieval

One of the open problems in content-based Image Retrieval is the recognition of material present in an image. Knowledge about the set of materials present gives important semantic information about the scene under consideration. For example, detecting sand, sky, and water certainly classifies the image as beach.

We try to tackle the problem of material recognition in two stages. First, the material reflectance characterized by invariant colour properties distinguish matte materials from glossy ones like metals. Comparison of the spatial response of various invariants leads to such a characterization. Secondly, the touch of roughness of a material may be charcterized investigating physical invariant texture properties. Therefore, we study the propagation of transformation groups through the Gauussian NJet. We demonstrate the Njet to charcterize the image as points in a high-dimensional scatter plot of the NJet components. Characterization of materials is then based on point cloud matching with prototype NJets. Finally, matching can be improved by fitting the NJet cloud to a statistical distibution. We show the image derivatives to obey a symmetric Weibull distribution, where the shape parameter varies between an exponential distribution and a Gaussian distribution. Matching the parameters of the Weibull distribution may lead to material recognition.