Descriptive Image Gradient from Edge-Weighted Image Graph and Random Forests

Creating an image gradient is a transformation process that aims to enhance desirable properties of an image, whilst leaving aside noise and non-descriptive characteristics. Many algorithms in image processing rely on a good image gradient to perform properly on tasks such as edge detection and segmentation. In this work, we propose a novel method to create a very descriptive image gradient using edge-weighted graphs as a structured input for the random forest algorithm. On the one side, the spatial connectivity of the image pixels gives us a structured representation of a grid graph, creating a particular transformed space close to the spatial domain of the images, but strengthened with relational aspects. On the other side, random forest is a fast, simple and scalable machine learning method, suited to work with high-dimensional and small samples of data. The local variation representation of the edge-weighted graph, aggregated with the random forest implicit regularization process, serves as a gradient operator delimited by the graph adjacency relation in which noises are mitigated and desirable characteristics reinforced. In this work, we discuss the graph structure, machine learning on graphs and the random forest operating on graphs for image processing. We tested the created gradients on the hierarchical watershed algorithm, a segmentation method that is dependent on the input gradient. The segmentation results obtained from the proposed method demonstrated to be superior compared to other popular gradients methods.