ELM based signature for texture classification

Texture classification is a very important field of image analysis. Because of this, the literature has been presented methods that provide more and more highly discriminative texture descriptors. This paper aims to contribute to the literature presenting a novel and powerful method to extract texture signature, based on weights of a single-hidden layer neural network called "Extreme Learning Machine" (ELM). We evaluate these descriptors in a classification experiment using three different texture datasets (and their rotated versions): Brodatz, Outex and Vistex. The results demonstrate a high classification accuracy of the method, proving that this novel methodology can be used successfully in computer vision problems. HighlightsIn this study, we model the texture as an Extreme Learning Machine (ELM).ELM is a single-hidden layer feed forward neural network with a very fast learning algorithm.We divided the image into small windows to compute the input and label of the ELM.Pixels of each window are the input, while the central pixel is the label of the ELM.We use the output weights of the ELM as a feature vector for texture classification.