Iterative Super Resolution

Instead of implementing the matrix-based algorithm as described in [MEAF97], I opted for a simpler iterative approach. The matrix approach requires knowledge of three entities:

• F: A warp matrix
• G: A blur matrix
• E: A noise vector

(verify this)

The iterative approach is simpler, and makes the following assumptions:

• Any frame can be obtained by applying a Euclidian geometrical transform to the frame of reference.
• There is no noise (to be addressed later, possibly using non-linear methods).
• There is no blur.

This allows us to create a very simple super-resoltuion algorithm. For every pixel in the reference frame:

1. "Ask" each of the other frames: what should the value of this pixel be? This requires the use of an Interpolation algorithm, since we often ask for the value of a pixel that is not actually present in the frame.
2. Use some technique (average, median filter etc.) to choose / calculate the most likely value of the target pixel.
3. Assign this value to the target image.

References

[MEAF97] Michael Elad and Arie Feuer. Restoration of a Single Superresolution Image from Several Blurred, Noisy, and Undersampled Measured Images. IEEE Transactions on Image Processing, 6(12):1646—1658, 12/1997.

---

[ Front | phdwiki | Index ]

Updated September 25, 2008

---