By the use of an epsilon filter, in the case where a plurality of different illuminations exist, a boundary between the illuminations can be appropriately extracted, and an unnatural image pattern can be prevented from being generated, therefore subjectively preferable compression of a dynamic range can be achieved. An edge strength G(x, y) is calculated per position on an input image, and a threshold E(x, y) of an epsilon filter (12) is controlled on the basis of the edge strength G(x, y). The epsilon filter (12) filters the input image on the basis of the controlled threshold E(x, y). On the basis of the edge strength G(x, y), the threshold E of the epsilon filter (12) is adaptively changed according to a local gradient of a pixel value I(x, y), so in the case of using a linear lowpass filter or a fixed threshold epsilon filter, an illumination boundary can be more accurately extracted.