An image processing apparatus includes an extraction unit that extracts, from among a plurality of designated regions in which labels of classes are designated, complicated regions which are regions of at least a part of the designated regions and are regions having relatively complicated contours, in an annotation image given as learning data to a machine learning model for performing semantic segmentation in which a plurality of the classes in an image are discriminated on a per-pixel basis, and a setting unit that sets additional labels for the complicated regions separately from original labels originally designated for the annotation image.