Complex Texture Features Learned by Applying Randomized Neural Network on Graphs

Since the 1960s, texture has become one of the most-studied visual attribute of images for analysis and classification tasks. Among many different approaches such as statistical, spectral, structural and model-based, there are also methods that rely on analyzing the image complexity and learning techniques. These recent approaches are receiving attention for its promising results in the past few years. This paper proposes a method that combines complex networks and randomized neural networks. In the proposed approach, the texture image is modeled as a complex network, and the information measures obtained from the topological properties of the network are then used to train the RNN in order to learn a representation of the modeled image. Our proposal has proven to perform well in comparison to other literature approaches in two different texture databases. Our method also achieved a high performance in a very challenging biological problem of plant species recognition. Thus, the method is a promising option for different tasks of image analysis.