A three-dimensional image without luminance irregularity is generated while achieving good contrast. An image processing apparatus is provided, including an image combining portion that generates combined images by combining, for each depth position in a specimen, a plurality of fluorescence images captured with differing exposure levels at each of different depth positions of the specimen; a smoothed-luminance calculating portion that calculates a representative luminance from the individual combined images and that calculates a smoothed luminance for the individual combined images by smoothing the calculated representative luminance in the depth direction; a luminance correcting portion that generates corrected images by correcting the luminance of the individual combined images on the basis of differences between the smoothed luminance and the representative luminance calculated; and a three-dimensional image generating portion that generates a three-dimensional image of the specimen from the plurality of corrected images.