PSYCH 221 FINAL PROJECT

We embarked on reducing the visibility of these artifacts. Fairchild asked us to best fill in these two rows so an observer would not  be able to notice the artifact. They noted that computation time would be important factor, expecting it to fall below 2-3 seconds. We decided to be open-minded and approach the problem with many different approaches. Also, we decided to measure the success of the results using several different measurements from S-CIELAB to L1 & L2 Norm. We will explain our approaches to the problem, then the image metrics used, and finally show you the results and our conclusions.

We came up with 4 different methods:

To evaluate the performance of each method, we used three metrics:

Since our project estimates two rows of pixels that are never captured, it is not possible to have tests that can determine the performance of algorithms on the missing rows with full confidence. However, a good indicator of how an algorithm would perform on the missing rows is how it performs on other rows of the image. Designating other rows as good as proxies for the missing rows would work well only if signal-to-noise-ration (SNR) is roughly uniform throughout the image. Given device specifications, this is a fair assumption to make. Furthermore, we chose the multiple testing rows randomly throughout the image (except for at the very top and the bottom), not necessarily very near the missing rows. This ensured that the content in the missing lines varied vastly between different testing rows, hence gave a good estimate of how the algorithms would perform on average. 

Given well-defined goals of minimizing visual artifacts and keeping computation time to 2-3s, we chose four metrics that measure to measure different algorithms' performance. S-CIELAB, L2 Norm and L1 Norm are metrics for measuring the how much corrected images varied from the original images. On the other hand, the computation metric simply measured the time an algorithm took to complete, hence served as an indicator of its computational complexity.