This work explores a statistical basis for a process often described in computer vision: image segmentation by region merging following a particular order in the choice of regions. We exhibit a particular blend of algorithmics and statistics whose error is, as we formally show, close to the best possible. This approach can be approximated in a very fast segmentation algorithm for processing images described using most common numerical feature spaces. Simple modifications of the algorithm allow to cope with occlusions and/or hard noise levels. Experiments on grey-level and color images, obtained with a short C-code, display the quality of the segmentations obtained.