Graph-based image gradients aggregated with random forests

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