Texture characterization via deterministic walks' direction histogram applied to a complex network-based image transformation

Texture classification involves acquiring descriptive features from the image. This work proposes a descriptor based on statistics from a complex network inspired transformation of the texture. The descriptor is generated by performing a deterministic walks algorithm on the image transformation, focusing on the representation of the shape of the walks to build the feature vector. The first innovation of the proposed approach involves creating a complex network from an image and performing walks using the values of the network of node degrees, instead of on the intensity of the original image's pixels. The second meaningful improvement is in the information that is obtained from the walks: instead of walk sizes or demanding fractal dimension computations, the proposed method derives shape information in the form of a walk direction histogram. Experiments applying the method for texture classification on several widespread data sets show that the proposed method improves correct classification rates compared to other state-of-the-art methods while using a smaller feature vector.