Project 1

Brief Overview

The project aims to restore the color images automatically from digitized Prokudin-Gorskii glass plate images. To achieve this goal, I extracted the three color iamge, stack them together after the automatic alignment so as to form a color image. The webpage's template is downloaded from w3.css

I implemented the auto-alignment method based on Sum of Squared Differences (SSD) (after normalization (minus mean and then divided by std)) with both single-scale implementation and pyramid version.

For the single-scale alignment, I align b channel with r and g channel respectively. Before performing the alignment, I first crop the r and g channel, so that I can search for which part of b channel matches r(g)'s cropped part the best. Then I will calculate the offset and then do the alignment.

For the pyramid version, I use skimage.transform.resize to downscale the original image by 8, 4, 2, and 1, and perform single-scale alignment on each level. The alignment range is [-int(cropped_x_length/10), int(cropped_x_length/10)], [-int(cropped_y_length/10), int(cropped_y_length/10)] for x and y axes on the lowest level, and [-2, 2] for the higher levels.

The problem I came up with is that if I use small window size for searching ([30,30]), then the program will not perform well on some of the images. Thus, I developed two strategies.

Approach 1: The first is to use the whole image for searching at coarsest level of the pyramid, and then use windows (like size of [30,30]) to search.

Approach 2: Another is to use the whole image for searching at all the pyramid levels. Since, the exhaustive search is down at a very low level, the latter computation cost is affordable (I only search for the range [-2,2], but it will definitely be much slower than the [30,30] approach). I will compare the performance of these two approaches in the Example Images sections.