A major challenge in computer analysis of complex scenes is the construction of an efficient recognition system for primary hypotheses generation in order to access large model bases. This problem, restricted here to man-made objects, presents two major issues. First, a perfect segmentation of the image is virtually impossible to obtain. Second, image features originating from a target object are hard to distinguish from those due to the background. This thesis proposes a new approach to this dual problem by delaying image features at different times according to their likelihood of being relevant fer recognition. In this way, the features selected in the early stages of the recognition process provide hints about the contents of the scene. Featuires which generate conflicting hypotheses are temporarily inhibited, which leaves time for confirming early hypotheses by actively searching compatible features. Once the inhibition delay has expired, previously conflicting features are taken into account. In this way, the system has the possibility to recover from erroneous interpretations. Results obtained on images of objects on non-uniform background are presented, that confirm the power of the approach.