Assignment #1 - Colorizing the Prokudin-Gorskii Photo Collection

Yu-Chen Lin

I. Introduction

Brief Overview

In this homework, we aim to align the RGB channels together to get a reasonable image.

My Approach to the Low-res Images

For the cathedral image, I conduct the following steps.

  • Noticing that the boundary may contain some unreasonable noise, I removed 10% of the pixels from each of the four boundaries.
  • In the lecture, the professor mentioned that the different brightness may cause some unreasonable comparisons. Therefore, I transformed the image to become zero-mean.
  • I preset the image translation range to be [-15, 15] times [-15, 15].
  • For each translation configuration, I use SSD loss to model the similarity and choose the best configuration with the least SSD loss. The image translated by the best configuration is then returned.

    • My Approach to the Large Images

    For the other larger TIF images, I conduct the following steps.

  • To avoid the high computational cost on the high-dimensional space, I downsample the image by a scale of 10.5%. I also do some parameter searches for the scale parameter by manually looking at the aligned results. The anti_aliasing parameter is set to True to avoid the aliasing effect during downsampling.
  • Align the channels in the smaller space using the algorithm from the cathedral image with the same parameters.
  • Scale the image back to the original size.

    • Challenges

  • Picking a good set of hyper-parameters is not easy. I try 10-20 values on the scale parameter and finally decide to use 0.105.
  • emir and self_portrait are the ones that do not align very well during my selection process of hyper-parameters.

    • II. Results

    G: (5, 2)
    R: (12, 3)
    Cathedral
    G: (5, 3) in the downsampled space
    R: (9, 5) in the downsampled space
    Emir
    G: (6, 2) in the downsampled space
    R: (13, 1) in the downsampled space
    Harvesters
    G: (4, 2) in the downsampled space
    R: (9, 2) in the downsampled space
    Icon
    G: (5, 1) in the downsampled space
    R: (11, 1) in the downsampled space
    Lady
    G: (8, 3) in the downsampled space
    R: (15, 3) in the downsampled space
    Self Portrait
    G: (5, 1) in the downsampled space
    R: (12, 1) in the downsampled space
    Three Generations
    G: (5, 0) in the downsampled space
    R: (9, 3) in the downsampled space
    Train
    G: (6, 2) in the downsampled space
    R: (12, 3) in the downsampled space
    Three Generations
    G: (6, 1) in the downsampled space
    R: (14, 2) in the downsampled space
    Village

    III. Prokudin-Gorskii Collection

    G: (2, 2) in the downsampled space
    R: (6, 2) in the downsampled space
    Altar
    G: (2, 0) in the downsampled space
    R: (6, -1) in the downsampled space
    Epiphany Church
    G: (2, 3) in the downsampled space
    R: (3, 3) in the downsampled space
    Savior Church
    G: (5, -1) in the downsampled space
    R: (10, -3) in the downsampled space
    Siren
    G: (0, 2) in the downsampled space
    R: (0, 15) in the downsampled space
    Sunset

    IV. Explanation

    My aligned results for the Emir and the Self Portrait ones are not too ideal. The reason may be that the same scale parameter may only work for some specific images. If you scale the image to be too small, once you translate the image, it may contribute to a large step in the original image space. On the other hand, if you scale down the image only a little bit, you may end up with only a small range of translations.

    Emir
    Self Portrait