The usual practice in pattern recognition is first to adopt some working definition for the patterns of interest and then to expend most of the effort on the equipment and/or automated techniques required to find and recognize these patterns. This paper is concerned only with the first step; it outlines a strategy, based on psycho-physical methodology, for determining those features of the pattern which are most likely to be informative. The approach used is termed psychopictorics; this is defined as a subfield of psychophysics which is concerned with pictorial stimuli, and in which it is assumed that, in a picture, the information of significance to the human observer may be characterized and analyzed in terms of the properties of perceived objects. The analysis of these properties involves the measurement of many psychophysical variables while the observer is responding to repeated, controlled changes of the features of single objects in the picture. Thus psychopictorics is strongly dependent on the development of computer picture processing techniques which permit such controlled manipulations without unduly degrading the quality of the picture.