Project 1: Color Alignment
Problem
This project addresses the problem of automatically aligning a series of photographs that differ spatially by some translational offset. Each photograph captures the same imagery, though different exposure settings produce irregular variations across the set of photographs, making this problem more challenging.The data used in this project comes from the Prokudin-Gorskii collection .
Algorithm
To solve this problem, I implemented both a single scale and multi-scale alignment algorithm, using a normalized cross-correlation metric to score image similarity. In the single-scale implementation, each image to be aligned was shifted over a window of translation offsets: [-25,25]x[-25,25]. While this approach produced nice-looking results, it took ~8 seconds to align three relativey small images (~300x400px).The multi-scale approach relied on producing a Gaussian pyramid of images, in order to perform mult-level alignment where approximate offsets are computed in low-res versions of the photographs, and subsequently refined in higher-res versions. The pyramid was created by continually downsampling the original images by a factor of 1/2, until the largest dimension of the smallest downsampled image was less than 40 pixels. In my implementation, approximate offsets were searched for over a 9x9 window ([-3,3]x[-3,3]). This produced correct-looking alignments in significantly less time (~0.8 seconds for the 300x400px images).
Extra credit
For extra credit components, I implemented automatic border-cropping, automatic white balance, and automatic contrast enhancement.My white balance algorithm follows the Retinex model, where only the red and blue channels are adjusted, taking into account what is known of human visual perception.
The automatic contrast enhancement technique I implemented is known as contrast-limited adaptive histogram equalization. As opposed to global histogram equalization, this method operates on local regions of the image to produce sharp contrast.