Here we use gradient domain fusion to reconstruct the original image. On the left is the original image and the right is the reconstructed image by solving the least squares objectives.
In this portion of the project we explore the concept of poisson blending within gradient domain processing. Gradient domain fusion has several applications, including seamlessly blending images together which we implement here. The goal is essentially to take a target image which we then want to smoothly place onto our source image (the background image) in the least noticeable way possible.
Gradient domain blending, but more specifically poisson blending, is able to achieve this by ignoring overall intensity in order to preserve the gradient of the source image so that changes to the source background after the target is placed is well matched. We are able to do this because the human eye can better adapt to changes in brightness without noticing anything overly different about an image. This can be set up as a least squares problem where we solve for the new intensities of the reconstructed image by summing the gradient values computed given the intensities of the target and source.
We must solve the following least squares objective:
Here v is the constructed image whose intensities we are solving for by solving the least squares problem. S refers to the source image and t refers to the target where i, j are the indices of the intensities of the source andt target.
Here we use gradient domain fusion to reconstruct the original image. On the left is the original image and the right is the reconstructed image by solving the least squares objectives.
Tupac at TedX is my favorite poisson blend result. I did not do anything specifically to ensure that the blending was particularly smooth for this image. However, the target background was mostly black as was the source background, so the result is most likely blended particularly well for this reason.
Pictured here is my failed result. Rather than being poorly blended, the dinosaur here just appears black. While at first I thought this was because my mask was incorrect, but it turns out the size of the dinosaur did not align properly with the target image.