The Classic Computer Vision Trick Behind Smooth Image Blending

This audio is presented by Hacker Noon, where anyone can learn anything about any technology.

The classic computer vision trick behind smooth image blending.

By Farzone Lotfi. If you have ever tried to create a multi-panel collage by stitching

many images together you have likely run into the problem of harsh, visible seams.

Slightly better would be some naive alpha blending. Slightly worse are simple cut and paste jobs.

Both leave unnatural transitions or ghosting artifacts. To get those buttery smooth,

seamless transitions, we can use a classic computer vision technique,

Laplacian Pyramid blending. Below you can see how awesome the results can be.

Building the pyramids. To blend two images, let's call them image A and image B.

Without harsh seams, we have to treat the broad strokes.

Low frequencies, like lighting and color, differently than the fine details,

high frequencies, like sharp edges. We doth this using image pyramids.

The Gaussian Pyramid, reduce. We take our original image and run a Gaussian filter, blur,

over it, then downsample it, e.g. shrink in 8x8 image to a 4x4, then 2x2, then 1x1.

Each step is a new, level, in our Pyramid, representing progressively coarser,

lower frequency data. The Laplacian Pyramid, expand. This isolates the details.

We take a smaller, coarser level of the Gaussian Pyramid and expand it to match the size

of the level below it. Because expanding is basically guessing the pixels in between,

it's not perfect. We subtract this expanded image from the actual image at that level to get

an error image. This error image is the Laplacian. It basically looks like an edge map containing

all the high frequency details. The blending process suppose you have image A, image B,

and a mask, region R, that tells the computer where the blend should happen.

In the process I learned the blending works from coarse to fine, 1.

Build the Laplacian Pyramids for image A and image B, L underscore A and L underscore B.

2. Build a Gaussian Pyramid for your mask, G underscore R. Blurring the mask as it gets smaller

ensures that the transition zone widens for lower frequencies. 3. Form a combined Laplacian Pyramid,

dollar L underscore O dollar, using the mosques Gaussian Pyramid as weights.

The formula at each level looks like this. 4. 5. Finally,

collapse the combined Pyramid, L underscore O, by continually expanding and adding the levels

back together to get the final blended image. The code IMPL EMEN TATI ONI wrote a Python

implementation that handles this exact mathematical process. It computes the Pyramids,

applies the blending step at each level, and reconstructs the output.

You can check out the full source code on my GitHub here. Laplacian blending with Python.

Or check out the colab directly. Conclusion BY breaking images down into their frequency bands

using Gaussian and Laplacian Pyramids. We can blend the structural details and the broad

lighting separately. The result is a composite image that fools the human eye into seeing

buttersmuth transition. Grab the code, create some image masks, and try making your own seamless

collages. Credit where it's due, computational photography course at Georgia Techie have

to mention that a fantastic resource for learning these concepts is Georgia Tech's CS 6475

Computational Photography Class. I took this course when I was in grad school and it was one of my

favorite courses. If you want a deep dive into the math and theory, I highly recommend checking

out these excellent course notes compiled by Monser Sala. The theory above leans on the foundational

concepts taught in that course. Thank you for listening to this Hackernoon story,

read by artificial intelligence. Visit Hackernoon.com to read, write, learn and publish.