ECE 563 Assignment 5

Images and m-files for this assignment may be downloaded from impro5.zip and WashingtonDC.zip.

Read Section 11.5 (p 675-683) on principal components. See also the routine princomp.m from DIPUM.

Individual Work

  1. Use your digitized color photograph and find its principal component transformation. Show the resulting transformed images (scaled), their eigenvalues, and eigenvectors.

Group Projects

  1. Correct the spatial variations in the image uneven.tif by fitting z = a0 + a1 x + a2 y + a3 rsq where x is a x-ramp image (0-1), y is a y-ramp image (0-1), and rsq is a quadratic ramp rsq = x.*x+y.*y. Determine the coefficients, subtract the fitted image from the test image, display the scaled difference image, and calculate the rms residual difference.

    Reference: Bernd Jahne, Digital Image Processing, 4th Ed., Springer, 1997. p. 229.

  2. Repeat the fit described above on the image c_backgr.tif. Divide the fitted image point-by-point into the image c_inhomo.tif. Threshold the resulting image to allow bwlabel to count the number of dark and light dust particles.

    Reference: Bernd Jahne, Digital Image Processing, 4th Ed., Springer, 1997. p. 236.

  3. Do problems 6.1, 6.3, 6.5, and 6.7. Note that the answer to starred problem 6.2 may help in doing problem 6.3.

  4. Plot your measured RGB values for the Macbeth chart images with the "true" values in the speadsheet. Find a linear transformation (RGB_corrected = C * (RGB_uncorrected) to perform a color correction. Calculate the rms difference for the color squares before and after correction. Note that there are two images of the Macbeth chart, one with a digital camera and one with a scanner. Which device had better color correction?

  5. Find the principal components of the Landsat multispectral images of Washington DC. Show the resulting transformed images (scaled), their eigenvalues, and eigenvectors. Compare the method in princomp.m to that discussed in class.

  6. Use the first three bands of the Washington DC image set to generate a "true" color RGB image.

Maintained by John Loomis, last updated 7 Feb 2005