To provide the three-dimensional shape recognition device which can accurately grasp the three-dimensional shape of a body by performing parallax extraction with higher accuracy.

This device is composed of a parallax image input device 11, an edge extraction part 12, an edge distribution calculation part, an edge correspondence determination part 13, a correspondence interpolation part 14, and a three-dimensional shape output part 15. The edge distribution calculation part calculates variation rates at edge points and edge intervals and the edge correspondence determination part 13 is composed of a preprocessing part equipped with a means which converts calculation information into similarity and a mutual connection type neural net consisting of connections of suppressibility and irritability and has a function for determining right-left correspondence relation between edges. The correspondence interpolation part 14 is composed of a similar mutual connection type neural net and has a function for determining right-left relation between points other than the edges, and the three-dimensional shape output part 15 calculates parallax and outputs a three-dimensional shape according to the calculated parallax quantity.