Graph-based image gradients aggregated with random forests

International audience; Gradient methods subject images to a series of operations to enhance some characteristics and facilitate image analysis, usually the contours of large objects. We argue that a gradient must show other characteristics, such as minor components and large uniform regions, particularly for the image segmentation task where subjective concepts such as region coherence and similarity are hard to interpret from the pixel information. This work extends the formalism of a previously proposed graph-based image gradient method that uses edge-weighted graphs aggregated with Random Forest (RF) to create descriptive gradients. We aim to explore more extensive input image areas and make changes driven by the RF mechanics. We evaluated the proposals on the edge and segmentation tasks, analyzing the gradient characteristics that most impacted the final segmentation. The experiments indicated that sharp thick contours are crucial, whereas fuzzy maps yielded the worst results even when created from deep methods with more precise edge maps. Also, we analyzed how uniform regions and small details impacted the final segmentation. Statistical analysis on the segmentation task demonstrated that the gradients created by the proposed are significantly better than most of the best edge maps methods and validated our original choices of attributes.