CS180 Images of the Russian Empire

Unaligned Cathedral

I initially divided the image into three, each representing the blue, green, and red channels respectively (from top to bottom).

After dividing them, I sought to align them and produce a color photograph.

Initial Approach - Simple Stack

First, I tried to see what would happen if I just divided the image in thirds, and stack them on top of each other.

The result, clearly, was not pretty.

Alternative Approach

Seeing this simple approach fail, I sought a different approach.

First, I used the circular_shift function to align the channels. This method shifts an image by a given offset along the vertical and horizontal axes, and enables us to test different alignments.

To measure the alignment quality quantitatively, I used the compute_ssd function to calculate sum of squared differences between the reference channel and the shifted target channel. Lower SSD values imply better alignment.

Then, I use the align_channels method to carry out a brute-fore search over a given range of shifts performed on the target channel. For every shift in this range, I calculate the SSD, and eventually choose the shift with the lowest SSD value.

For large images (i.e. the .tif files), a brute force becomes super expensive and would take a long time to run. To address this, I implemented the pyramid_align method. In this method, we create an image pyramid that is made up of progressively smaller images, each downsampled from the previous one by a factor of 2. The pyramid method aligns the images at the smallest scale, where the search space is minimal. I then refine this shift at higher resolutions by upscaling the shift.

To build this image pyramid manually, I've written the manual_downsample and manual_upsample methods, which decreases/increases the image size.

Finally, the proess_image method carries out this entire process. It reads the image, divides it into three channels, and crops each channel to remove edges (default factor of 5%). The green and red channels are aligned to the blue channel using the pyramid alignment method. Finally, I stack these aligned channels to form the output image, which is saved and displayed.