Abstract : In computer vision, the way the problem is addressed is very important. As regards the detection of the first visual features, like edges or regions, the theoretical approach is difficult because the goal is not clearly defined. Our approach consists in building step by step a satisfactory solution, essentially taking into account the way information has to be processed. We propose simple principles to get an efficient information management and powerful decisions strategies. An incremental control structure seems to be a very efficient way to apply our principles and to prepare an heuristic- based approach. For example, by translating our expertise in an algorithm, we have built a new edge detector with adaptive thresholding, which results appear to us better than those of classic edge detectors for a great number of images. This edge detector is unsupervised and detect "step edges" as well as "line edges". We also propose a cooperation between our edge detector and a region growing process, that agregates pixels one by one to build regions. Compared to other cooperating algorithms, our control structure is much more flexible, the information management allowing to get, at the right time, at the right place, the complementary information needed to take the right decision. However, we still have to make progress, using more information, in particular concerning the shape of the regions being built. Our conclusion is that segmentation suffers from a too restrictive modelisation. To carry on making progress in that domain, we may try to address problems differently.