Project 3: Gradient Domain

Methodology

The project is on gradient domain fusion. In every set below there is a source and a target image, with the goal being to seemlessly transfer a region in the source image to the target. In general, the goal is to create a fused image which matches the gradient of the source image within and up to the boundary of the selected region. In general, this is not possible, so I employ a least-squares approximation. All results are below.

Results

Source	Target	Naive Transfer	Blended Transfer	Mixed Gradient Transfer

Analysis

The first two examples above are easy cases where both the source and target are relatively low-frequency on the boundary. The method works fairly well, and single vs mixed gradient makes no difference.

The next example (face on coin) shows a low-frequency source on a high-frequency target. In the blended transfer, the edge is noticable because of a sudden change in high frequency content. In the mixed gradient, the target seems to overpower the source and disrupt the signal.

Contrast this to the next example, which has a high-frequency source and target. Here, both blending operations work fairly well, since artifacts are difficult to distinguish from the natural noise.

The next two examples are experiments in watermarking and logo superposition. In both cases, blended transfer is a bad idea, since the details of the background image are lost. Mixed gradient works fairly well, with particularly nice results for the Will Rice College logo on brick. If not for the slightly non-orthogonal view of the wall, it would be very difficult to tell that the logo is not painted on.