![Logo](/Boazius/TAU-Graphics-Ex1-SeamCarving/raw/master/images/tauLogo.jpg)
Content-aware image resizing using Seam Carving, created as part of the TAU Graphics course
Report Bug
·
Request Feature
Exercise 1, As part of the "Fundamentals of Computer Graphics, Vision and Image Processing" course in Tel aviv university (2022).
Made by Boaz Yakubov and Noga Kinor
Content-aware image resizing changes the image resolution while maintaining the aspect ratio of important regions - This is different from simple stretching or shrinking of the image, since it intelligently avoids impacting "important" areas of the image, and instead modifies the least important areas of the image (typically the background)
For example, for an image like this (1920x1080):
if we wanted to resize it to 1500x1100 , we must find the "unimportant" areas - horizontal marked in black and vertical marked in red:
and then delete these unimportant areas in the image to get the final resized image
Note that the camels were not resized at all! only the background sky and desert was shrunk.
- the resized image
- the original image with horizontal black lines (that will be deleted or duplicated)
- the the partially resized image with vertical red lines (that will be deleted or duplicated)
The program optionally uses a "forward looking energy function" instead of the regular energy function, in order to reduce artifacts in the resized image
In order to implement the Seam carving algorithm, we need to define an energy function that specifies the "importance" of each image pixel. we can calculate the pixel importance using the image gradient.
We computed it using the formula
to find the optimal "Seams" to remove or duplicate , i.e the lines along the image:
- calculate the cost matrix M
- do k times:
- find a path with the lowest total energy from one end of the image to the other.
- delete or duplicate it
to resize the image both horizontally and vertically, we first change the width using the algorith and then the height on the partially resized image.
Seam carving can introduce artifacts to the resized images. to reduce these artifacts, we implement a forward-looking energy function:
$M(i,j) = E(i,j) + \min\begin{cases} M(i-1,j-1)+C_L(i,j), \ M(i-1,j)+C_V(i,j) \ M(i-1,j+1) + C_R(i,j) \end{cases} $
where (
when using the program, you can choose to use the previous energy function, or this forward looking energy function
To get a local copy up and running follow these simple steps.
- clone or download the repository
- open shell in the repository folder
- verify python and pip are installed ( use
python --version
andpip --version
) - install requirements using
pip install -r requirements.txt
- copy your desired images(s) into the folder, preferably into the imagesInput folder
now, to run the application simply type:
python main.py
with the following arguments
- image_path (str) - an absolute / relative path to the image you want to process
- output_dir (str)– The output directory where you will save your outputs.
- height (int) – the output image height
- width (int) – the output image width
- resize_method (str) – a string representing the resize method. Could be one of the following: [‘nearest_neighbor’,‘seam_carving’]
- use_forward_implementation – a boolean flag indicates if forward looking energy function is used or not.
- output_prefix (str) – an optional string which will be used as a prefix to the output files. If set, the output files names will start with the given prefix. For seam carving, we will output two images, the resized image, and visualization of the chosen seems. So if --output_prefix is set to “my_prefix” then the output will be my_prefix_resized.png and my_prefix_horizontal _seams.png, my_prefix_vertical_seams.png. If the prefix is not set, then we will chose “img” as a default prefix.
so for example
python main.py --image_path "imagesInput/tower.png" --output_dir "imagesOutput/" --height 900 --width 900 --resize_method "seam_carving" --output_prefix "my_prefix" --use_forward_implementation
will run the seam carving algorithm on tower.png and output the files:
- my_prefix_resized.png
- my_prefix_vertical_seams.png
- my_prefix_horizontal_seams.png
- Python
- pip
Distributed under the MIT License. See LICENSE.txt
for more information.
I can be reached at at my email: boazyakubov@gmail.com
- Noga Kinor for wonderful teamwork in this course
- Professor Daniel Cohen-Or for teaching this course at TAU
- Roey Eliyahu Bar-On for instructing us on the subject at TAU