In image thresholding problems, no clear answer has been given to the simple question of what the thresholding result should be. It is argued that image thresholding should produce a simple binary picture because a complicated picture is not easy for human perception. Three types of complexity measure for binary pictures are defined, and it is shown how they match the human perception of complexity. Then a minimal complexity measure thresholding method is proposed to get simple binary patterns from multivalued (or gray-scale) images. A hierarchical algorithm is also developed in which a depth-first picture expression (DF-expression) is used. Experimental studies have demonstrated the advantage of this method.