This paper analyzes and evaluates various methods for rotation of images. Three different algorithms, all stemming from different factorizations of the basic rotation matrix, are employed: cosφ{symbol} -sinφ{symbol} sinφ{symbol} cosφ{symbol} Factorized versions of the rotation matrix correspond to execution in more than one pass. Such executions require less fast memory (which might be important for rotation of very large images) and may also be faster. All rotation algorithms require interpolation. In the present paper bilinear interpolation has been employed as well as bicubic functions of size 2 × 2, 4 × 4 and 8 × 8. An image is rotated an angle θ{symbol} and this result is then rotated back to the original position. The magnitude of the differences with respect to the original is taken as an indicator of the quality of the rotation algorithm and the interpolation technique. For very high fidelity, a three-pass rotation seems to be advantageous. © 1992.
