We are glad to announce the release of MLDenoise, a new machine learning tool for noise reduction of astronomical deep-sky images. This is a preview of the next generation of machine-learning PixInsight tools that we'll be releasing in the coming weeks and months, including noise reduction, star removal, deconvolution, and other tasks where specialized neural networks can help solve intractable or ill-posed problems with hardware acceleration.
This release is for the macOS ARM64 (Apple Silicon) platform exclusively, and is available as a regular update for PixInsight 1.9.4 Lockhart, builds 1693 and later. We'll release a new version of MLDenoise for Linux, macOS, and Windows in July, including many improvements and features. In this technology preview release, MLDenoise is fully functional but still lacks some important features, such as real-time preview capabilities. It works superbly well with previews, thanks to the implemented caching techniques that optimize repeated execution.
Machine Learning Model Files
To use MLDenoise, you need a specific neural network database file. Right now you can download the first one from our software distribution system:
Download the .xmlm file to any folder on your local filesystem (such as a folder under your home folder, for example). Then run MLDenoise and click the Preferences (wrench icon) button to open the MLDenoise Preferences dialog, where you can select the default model file.
Running MLDenoise
On macOS, MLDenoise will automatically use GPU acceleration via Apple's Core ML infrastructure. This is very convenient because hardware acceleration is available on macOS as an integral part of the operating system, without requiring any additional configuration.
We recommend using MLDenoise on color-calibrated, linear deep-sky images, where the result will be optimal. To accelerate the trial-and-error work, you can use reduced previews on selected image regions of interest. Once you have fine-tuned the required parameters—which is really easy with the implemented interface—you can apply the process to the entire image, as usual.
We have implemented a dynamic cache feature that stores partial results and other auxiliary data structures while you try different parameters on previews. This greatly accelerates the task by avoiding unnecessary duplicate calculations. The implemented dynamic cache is perfectly usable, but still not ideal. We'll improve this feature considerably in the next version of the MLDenoise.
Linear Masks
We have implemented linear masks in several tools for a long time, including MultiscaleLinearTransform and MultiscaleMedianTransform. Use a linear mask as a local support for noise reduction of linear images. Linear masks can be very efficient for noise reduction of linear images with a strong correlation between signal-to-noise ratio and illumination. This condition holds for most linear deep-sky astronomical images.
To evaluate the protection provided by a linear mask, turn on the Preview mask option. When you execute the process with this option enabled, the target view is replaced with a representation of the mask that will be applied during normal MLDenoise execution. Here is an example:
Note that after running MLDenoise with the Preview mask option enabled, the screen transfer function (STF) of the target view is disabled to allow you to view the generated mask. For normal execution on the same view, you'll have to re-enable STF by pressing Control+S on macOS (F12 on Linux and Windows).
This release is for the macOS ARM64 (Apple Silicon) platform exclusively, and is available as a regular update for PixInsight 1.9.4 Lockhart, builds 1693 and later. We'll release a new version of MLDenoise for Linux, macOS, and Windows in July, including many improvements and features. In this technology preview release, MLDenoise is fully functional but still lacks some important features, such as real-time preview capabilities. It works superbly well with previews, thanks to the implemented caching techniques that optimize repeated execution.
Machine Learning Model Files
To use MLDenoise, you need a specific neural network database file. Right now you can download the first one from our software distribution system:
Download the .xmlm file to any folder on your local filesystem (such as a folder under your home folder, for example). Then run MLDenoise and click the Preferences (wrench icon) button to open the MLDenoise Preferences dialog, where you can select the default model file.
Running MLDenoise
On macOS, MLDenoise will automatically use GPU acceleration via Apple's Core ML infrastructure. This is very convenient because hardware acceleration is available on macOS as an integral part of the operating system, without requiring any additional configuration.
We recommend using MLDenoise on color-calibrated, linear deep-sky images, where the result will be optimal. To accelerate the trial-and-error work, you can use reduced previews on selected image regions of interest. Once you have fine-tuned the required parameters—which is really easy with the implemented interface—you can apply the process to the entire image, as usual.
We have implemented a dynamic cache feature that stores partial results and other auxiliary data structures while you try different parameters on previews. This greatly accelerates the task by avoiding unnecessary duplicate calculations. The implemented dynamic cache is perfectly usable, but still not ideal. We'll improve this feature considerably in the next version of the MLDenoise.
Linear Masks
We have implemented linear masks in several tools for a long time, including MultiscaleLinearTransform and MultiscaleMedianTransform. Use a linear mask as a local support for noise reduction of linear images. Linear masks can be very efficient for noise reduction of linear images with a strong correlation between signal-to-noise ratio and illumination. This condition holds for most linear deep-sky astronomical images.
To evaluate the protection provided by a linear mask, turn on the Preview mask option. When you execute the process with this option enabled, the target view is replaced with a representation of the mask that will be applied during normal MLDenoise execution. Here is an example:
Note that after running MLDenoise with the Preview mask option enabled, the screen transfer function (STF) of the target view is disabled to allow you to view the generated mask. For normal execution on the same view, you'll have to re-enable STF by pressing Control+S on macOS (F12 on Linux and Windows).
