Fits Bezier surface over each channel of image where the height at (x,y) for channel c is the pixel value of the image at (x,y) for channel c.
- The motivation is to represent a digital image as a continuous (analog) image.
- Uses Keras with Tensorflow backend to fit Bezier surface of degree
(m,n)provided by user. - The learned control points (
(m+1)*(n+1)in total) can act as a low-dimensional representation of the image (if(m+1)*(n+1)<H*W).