PURPOSE: To realize the high-speed processing which determines a threshold for binarization by a neural network learnt based on values representing the histogram of lightness of a partial image of an input image.

CONSTITUTION: The image of a character picture is picked up by an image sensor 10 and is digitized by an A/D converter 11 and is stored in a frame memory 12. The character image is divided into partial image blocks of a prescribed size by a window circuit 13 and is successively read out thereafter. The histogram of lightness of a partial image is stored in a histogram memory 15. A maximum lightness and a minimum lightness are detected by a maximum and minimum detector 16, and the lightness of the maximum frequency and a proportion of the maximum frequency to the number of pixels of the partial image are detected by a peak detector 17, and a neural network 18 specifies the threshold in accordance with these detected values, and this threshold is used to perform binarization.