In this paper, an image compression method based on a novel idea of surface fitting is proposed. An image is divided into sub-images using a quad tree according to the complexity of the image surface, which is a three-dimensional surface when the pixel values are considered as heights (z-values) defined on the x-y plane. Sub-image surfaces are approximated using a radial-basis function network (RBFN), and the parameters of the RBFN are stored as compressed image representation. This method has comparable performance and quality to the JPEG image compression technique.