The present invention relates to an image processing device and method that enable generation of a highly precise prediction image using a small amount of control information. A motion compensation circuit specifies a macroblock corresponding to a prediction image in part of reference frames using a motion vector supplied from a prediction mode determination circuit, reads an image thereof from a frame memory, and extracts it as a motion compensation image. A motion prediction circuit reads, from a frame memory, at least one of the remaining reference frames, performs motion prediction of the motion compensation image MC0 supplied from the motion compensation circuit in each frame, reads an image of a macroblock that matches or is similar to the motion compensation image MC0 from the frame memory, and extracts it as a motion compensation image.