Noise robust and rotation invariant framework for texture analysis and classification

Texture feature extraction is an important task in image processing and computer vision. Typical applications include automated inspection, image retrieval or medical analysis. In this paper we propose a noise robust and rotation invariant approach to texture feature analysis and classification. The proposed framework is based on a simple texture feature extraction step which decomposes the image structures by means of a family of orientated filters. The output is an orientation density vector which is called Orientation Feature Vector (OFV). The OFV can be used as the input feature vector to a texture classifier, and in addition, by using an interpolation step it is possible to extract the main orientations of the texture from the OFD, providing additional high level features for image analysis. In this work, three families of filters have been studied in the texture feature extraction step. The experimental results show the ability of the proposed framework in classification problems, improving more than 20% the results of other state-of-the-art methods when a high level of Gaussian noise is considered.