At present, the procedure which is most widely used to identify an object consists of displacing its image in order that it be as close as possible to a reference pattern, and then to make a decision based upon the value of an Euclidean distance. This decision may be either in a deterministic framework or in a statistical one. The transformations so involved are translation and rotation. It is shown that, if one takes for granted the invariance of the amount of information involved by the object considered as an informational source, then there is another class of transformations, referred to as deformation, which should be also taken into account. The question is examined, and its consequences in the design of artificial vision are outlined in both the deterministic and stochastic cases.