Assignment 1

Grace Su (Andrew ID: gdsu)

1. Single-Scale Version

I first cropped the image by a fixed value on all sides to remove the border around the actual image. Then, to align 2 channels, for displacements in the range (-15, 15) in both the x and y directions, I calculated the sum of squared differences (SSD), and used the displacment with the smallest SSD for the final alignment.

My implementation calculated that the green channel had a (y,x) offset of (-8, 2) from the blue channel, and the red channel had an offset of (-14, 3) from the blue channel.

The output image is shown below:

2. Multi-Scale Version

I first cropped the image by a fixed value on all sides to remove the border around the actual image. Then, for my multi-scale alignment procedure, I used 5 image scales, where each successive scale was smaller by a factor of 2. I used my single-scale alignment function to calculate the best displacement at each scale. I then scaled the displacement by a factor of 2, then used that result as the initial displacement for the next scale.

2.1 Results on given example images

2.2 Results on other images from the collection

3. Bells and Whistles

I also added the following bells and whistles:

• Automatic contrasting: I did im = (im - np.min(im)) / (np.max(im) - np.min(im)) before separating the image's color channels.
• Better features: Inside my single-scale alignment function, I compared the SSD of the canny edges instead of the original pixel values.

Here is a comparison of before and after adding the bells and whistles: