ColorizeXL

Python implementation of the following paper by Anat Levin, Dani Lischinski, Yair Weiss, puublished in SIGGRAPH04.

We have adapted their technique to make it suitable for large images, by using overlapping patches with annotations.

Installation:

pip install colorizexl

Usage

Colorization of Black & White images

Previous to using the package, you have to setup the following:

• original black & white image
• annotation color image (please check details in the paper)

Recolorization of color images

Previous to using the package, you have to setup the following:

• original color image
• annotation image (please check details in the paper): white where the color stays the same, different color where you want to recolorize

Command line

Call the colorizexl.py from command line, to colorize or recolorize images.

python colorizexl.py --annotated=<path_to_annotated_image> --grayscale=<path_to_image_to_(re)colorize> --output=<path_to_(re)colorized_image> --recolorize=True/False

Python code

from colorizexl import ColorizeXL

colorizer = ColorizeXL(
        input_grayscale_name, input_annotation_name, recolorize=False
    )

output = colorizer.colorize(step_size=patch_size, overlap=overlap, n=neighbour)

Customization

From Python you have the options to finetune the following parameters:

• step_size (default 100), defines the size of the square patches
• overlap (default 2), defines the overlap of the patches
• n (default 10), defines the radius of the neighbours in a circle for computing the color

Package organization

colorizexl

Contains the ColorizeXL class, the engine behind the (re)colorization. Its functioning is as follows:

• __colorize: will take an image and an annotation as inputs and will colorize the image; it will return a colorized image
• colorize: main entry point in the algorithm, this method will split the initial image in patches and send them to __colorize. It will also take care of the overlaps to ensure smooth transition between colorized regions
• colorize_no_patch: wrapper for __colorize method, this function is the initial implementation of the algorithm, which is slow and will crash for large images

notebooks

Contains the notebooks used for experimenting:

• CS445_usage: example code for colorizing and recolorizing images (example images from the original paper are provided); output of the example images can be seen here
• CS445_master: our inital experiments with colorization
• CS445_patch_sort: experiments with intelligent sorting of patches to be recolorized

tests

Contains the test code used for fine-tuning and evaluating our package.

• test_hyperparameters: loops over the hyperparameters (patch size, overlap and neighbours) to compute the MSE (mean squared error), SSIM (structural similarity) and PSNR (peak signal to noise ratio); from these, we have mainly relied on MSE; it also saves an image for each combination, to allow manual selection of the most pleasing visual effect
• test_timing: downscales a large image by a factor of 0.1 to 1.0 (step 0.1) and measures the time needed for computation (please note that the images resulting from downscaling will not necessarily be of a high quality); execution time of our code scales linearly with the number of pixels to recolorize
• test_timing_no_patch: similar to test_timing but using a one-shot colorization; this original method presents an exponential increase in time per pixels to be colorized