Seam Carving in python, and it comes with a flask application! Original paper: Seam Carving for Content-Aware Image Resizing

Seam carving is an image processing algorithm for content-aware image resizing. Simply put, it can resize images without distorting important objects.

Demonstration:

image	original	simple resize	seam carving
demo1	(1428x968)	(968x968)	(968x968)
demo2	(800x533)	(400x533)	(400x533)
demo3	(1280x800)	(1000x800)	(1000x800)

Consider resizing the mages/Broadway_tower_edit.jpg from 1428x968 to 1427x968 for the moment. We first calculate a floating-point score for each pixel in the image. It represents the importance of that pixel as shown below. The brighter the pixel, the more important it is.

And then, all we have to do is to find a path or a seam from top to bottom which accumulates to the minimum score, and remove every pixel along that seam. In this case, it is the highlighted green seam.

Repeat this process a couple of times, we will be able to resize the image from 1428x968 to 968x968 without doing much damage to the content of the image.

As you might have noticed by now, the most important contribution of the seam carving paper is that it provides a way to evaluate the "importance" (called the energy) of each pixel in the image. And since the paper was published, many other evaluation algorithms have been proposed.

In this project, I implement the original method in the paper. It makes use of two Sobel filters, one for the verticle direction, and the other one for the horizontal direction. Together they show how much different a pixel is from its surrounding pixels. The dissimilar the pixel, the higher the energy score.

One obvious disadvantage of Seam Carving is that it runs insanely slow. On CPU with no optimization, it takes 42 seconds to process one image. That's where CUDA comes in. CUDA enables developers to run general-purpose programs on and harness the great parallel computing power of NVIDIA GPUs. With CUDA acceleration, I was able to reduce the processing time to 0.5 seconds! For more information about my hardware setup please refer to the Comparison of speed section.

# install python packages
python3 -m pip install -r requirements.txt
# see usage
python cml_ui.py -h
# resizing with seam carving
# the result image is save as res.jpg
python cml_ui.py -r 512 -c 700 images/rem.jpeg
# object removal with seam carving
python cml_ui.py image.jpg --protect mask_protect.jpg --remove mask_remove.jpg

And it comes in flask flavor!

python web_ui.py

and then go to http://127.0.0.1:10800/ in your favorite browser

I use Linux's time program to measure the time consumed. Each version is tested with the following command:

time python cml_ui.py images/rem.jpeg -r 400 -c 600

My hardware setup:

• CPU: AMD Ryzen 7 5800H (16-core)
• GPU: NVIDIA GeForce RTX 3080 Ti
• another GPU: NVIDIA GeForce RTX 3090

Results:

real    0m0.564s
user    0m0.506s
sys     0m1.246s

real    0m1.763s
user    0m1.702s
sys     0m4.742s

real    0m42.481s
user    0m45.133s
sys     0m4.929s