Fractal descriptors of texture images based on the triangular prism dimension (Q1716781)

In the paper the authors study the use of fractals in the description of images for the pattern recognition task. In their study they propose a new type of fractal dimension derived from the similarity dimension, which is named triangular prism dimension. It is based on the relation between the surface area of a triangular tessellation of a grey-level map and the area of the base of each triangle. To improve the ability of identifying multiscale patterns along the texture the authors introduce an exponent weight. Moreover, in order to reduce the dimensionality of the feature set they use the Karhunen-Loeve transform. After introducing the dimension, the authors give its mathematical interpretation which is based on the theory of fractional Brownian motion. To show the effectiveness of the proposed fractal descriptors, the authors present some experiments carried out on four standard data sets, i.e., Brodatz, Vistex, UIUC and KTH-TIPS2b. The experiments are divided into two groups depending on the task. The first group of experiments was made for the image retrieval task, whereas the second group consists of the image classification problem. Moreover, the authors verify the invariance of the proposed method under Gaussian noise and rotation by predefined angles. The last conduced experiment is related to the computation time of the method. The experiments show that the method proposed by the authors is more robust than all the other compared approaches, including state-of-the-art approaches.