CS 7321: Computer Vision I

• The Problem
• The Solution
• Assumptions and Weaknesses
• Possible Improvements
• The Results
• The Code

This problem set plays around with multi-resolution techniques for images.

1. Implement the Adelson and Burt algorithm for splining overlapping images. The input images can be of any size. This will involve writing programs to compute the Laplacian pyramid for a given color image and the Gaussian pyramid for image masks. Implement Expand and Reduce as the basic routines for constructing the pyramids. You can use a 5x5 kernel with the weights as given in the paper.
2. Use your program to spline a pair of color images of your choice.
3. Use your program to form a composite images using combinations of images from http://www-astronomy.mps.ohio-state.edu/~dixie/animals.html OR
4. Please feel free to gather images from the WWW or collect them via camera or scanner.
5. For extra credit, compute an additional montage for a sequence of moving video images.
6. Include a one page description of how your program works. Be sure to discuss the assumptions and weaknesses of the particular technique employed. What are the limitations of this image splining technique?
7. Include images showing the Gaussian and Laplacian pyramids for one of the two images.
8. The program source and your output "spliced" images.

1. Preparation of the images: convert them into three sub pictures in R/G/B formats for later use.
2. A covolution mask [5x5] is created.
3. Build reduce/expand funciton
   • Reduce: It consists of Gaussian-low-pass filter. Reduce the the size of the matrix.
   • Expand: subtract an Expanded version of each pyramid layer from the previous layer. Expand() consists of supersampling the former version to the next higher size, followed by filtering by the same Gaussian filter.
   • Build the Laplacian/Gaussian pyramids for the images to be merged. Recursively run the Reduce and Expand functions.
   • Combining the images together at each individual level, and recover the resulting Laplacian image to an actual image. (Sinlge color factor)
   • Combine the results above from three color subpictures to form the final result.

Assumptions that I made:

• Merged images are aquare and of same size.
• Their dimesions are of 2^N +1.

Weaknesses:

• Strict lemitation of the input images (shape/dimension).
• Splining is executed at the central border of them. And the left 50% of the left image and the right 50% of the right image are mixed. Only the pixels are the central line are averaged. The efffect of the splined is not very good sometimes.

• The size of the image could be more flexible, by store the row/col info by a better data structure.
• The splining area could be adjustable, or set a mask to locate a speticular area of one image into another.

• mul_spline.m - Splines left/right image
• Reduce.m - Reduce function.
• Expand.m - Expand function.
• Pyramids.m - Builds Gaussian or Laplacian pyramid.