The code makes use of the NPP utils library to combine an list of images in a folder (see below)
onto a a single output image. By doing so, one can aim to increase exposure, signal to noise ratio, or simply mosaic astronomical images into a single image.
cudamosaic has been tested with Ubuntu 22.04 and cuda 12.6 and above. We first suggest you install the cuda packages. Please see https://developer.nvidia.com/cuda-downloads for latest cuda installation.
Make sure that nvcc is also in the PATH directory
ubuntu@ip-10-255-9-77:~/cudamosaic$ which nvcc
/usr/local/cuda-12.8/bin//nvccIf not please add the cuda folder in your PATH environment variable.
export PATH=/usr/local/cuda-<version>/bin:${PATH}where version is in our case 12.8.
We also need access to opencv
sudo apt-get install libopencv-dev
Once this is done, we are ready to install cudamosaic
git clone https://github.com/FabienVoisin/cudamosaic.git
Then one needs to clone the NPP utils repo
cd cudamosaic
git submodule init
git submodule update
Then type
make
to compile the code
In order to test the code, one can use the Orion folder or testimage folders via the following command
./mosaic -d testimage -o testimagemosaic.png
The output image should in this case be testimagemosaic.png. The image finalresultexp.jpg also provides the final exposure map from aggregating all the input images.
There are currently two command line parameters:
-d (required): path of folders which has the input images
-o (optional): filename of the output mosaic image. In the case, this is not explicitely parsed, then defaultfilename.jpg would be used.
The code first creates a greyscale image of the first image. Based on a specific threshold, the code then highlight pixels above certain values. It will then select a box of the signals around the pixel with the highest intensity. The code will use this box to perform correlation maps with other images. We expect the maximum value of the correlation maps will determine the offset position between the new input images and the mosaic images. Once that offset position is calculated, the code add the values of the new input images onto the mosaic output image. The final output image is then normalised by dividing the output image by the exposure map.
In the future, all input images should be processed independently in streams.
The code currently does not provide any method to measure how well the correlation signal box managed to match with the input images. A threshold score would reject input images with potentially incorrect position.
In some case, the image may need rotation as well as translation to perfectly match with the reference image.
The algorithm currently picks the pixel with the highest grey-scale intensity value to create a box of signals. However, this would work best to choose a box with the least amount of degeneracies (e.g permutations in signals within the box that will lead to the same result) to avoid incorrect offset position.