Awesome
"PBR" filter
{P}ansharpening by {B}ackground {R}emoval algorithm for sharpening RGB images
By Daniel Buscombe
Sample imagery comes from the Aeroscapes dataset and processed with the PBR filter using default settings
<!-- ![](https://github.com/dbuscombe-usgs/PBR_filter/releases/download/0.0.0/short_small.gif) -->- Read image, wavelet denoise, and convert to HSV
- Do 'inverted background subtraction' on the V (intensity) channel
- combine with HS, convert to RGB
The effect is to sharpen details of object boundaries/transitions, brighten, and recolour. But, without changing the overall distribution of values within the image (i.e. the range), and recolouring in an internally consistent (deterministic) way
Sample 'Madeira' imagery comes from this USGS data release by Brown et al
Sample 'OBX' imagery comes from this USGS data release by Ritchie et al
- Create a conda environment
A. Conda housekeeping
conda clean --all
conda update -n base -c defaults conda
B. Create new pbr
conda environment
We'll create a new conda environment and install packages into it from conda-forge
conda env create -f install/pbr.yml
activate:
conda activate pbr
- Run the program
python PBR_filter.py
To get thicker dark lines use
python PBR_filter.py -r 5
To get even thicker dark lines use something like
python PBR_filter.py -r 7
If you want to print out two more plots showing the process step-by-step for each image, use :
python PBR_filter.py -p 1
navigate to a folder of images you wish to filter, and it will step through them one by one
- Visualise your results see the png images the program makes inside the same directory as the input images
How does this work?
Let's take this image, which is a tile cropped out of a larger orthomosaic - see here:
The filter with default radius of 3 pixels creates this PBR image
In the figure below, the process is broken into stages
a) original image
b) wavelet denoised image, where noise over a range of scales is removed and mostly affects very small scale (pixel level) noise. This step isnt crucial but I always like to denoise imagery if I can as a precaution
c) the greyscale background image that has been created with a rolling ball filter with ball of radius [whatever]
d) greyscale version of the denoised image
e) the intensity image that is the greyscale divided by the greyscale background image. This image accentuates edges and makes intervening areas almost uniformly bright
f) the filtered RGB image that is the result of swapping the greyscale with the intensity image in the HSV stack of the original RGB image, then converting that into RGB colorspace
Here's another example (from a 1m NAIP image)
Original image of a saltmarsh environment
The filter with default radius of 3 pixels creates this PBR image
In the figure below, the process is broken into stages
disclaimer: I do not know if I have reinvented the wheel - I have not searched for similar implementations. Please tell me by opening an Issue if this technique has previously been proposed