Pixels in a window, which corresponds to the position of an input pixel and has a predetermined size, are extracted from a reduced image obtained by reducing an input image to a predetermined scale. Substitute data used to substitute the value of the input pixel is generated based on the extracted pixels in the window. The difference value between the substitute data and input pixel value is calculated and is compared with a first threshold value. When the difference value is smaller than the first threshold value, the input pixel value is substituted by the substitute data. Thus, low-frequency noise reduction can be attained using a minimum required memory size while suppressing adverse effects such as a resolution drop and the like.