Feature points (41, 42, 43) in the heat image (10) of a casting die (1) are extracted and a predetermined geometrical conversion processing is performed on the heat image (10) such that the feature points are superimposed on the reference feature points (61, 62, 63) set in a reference heat image (30) picked up previously to generate a corrected heat image (20). A difference image (40) is generated by superimposing the corrected heat image (20) and the reference heat image (30) such that the corrected feature points (51, 52, 53) in the corrected heat image (20) is superimposed on the corresponding reference feature points (61, 62, 63). With such an arrangement, a highly reliable difference image can be generated even when the imaging field of vision slips off among a plurality of heat images.