DeepSNR - Noise Reduction Tool

Overview

DeepSNR is a deep-learning-based denoising tool for high-quality astrophotography images. It is designed to remove noise while preserving fine details, making it ideal for processing raw integration results.

⚠ Best used immediately after image integration. Avoid using it on already processed images, as it assumes uncorrelated noise.

Parameters

Strength

Controls how much noise is removed by blending the original and denoised images using a linear combination. I advise against using this slider: run at full strength and experiment with the ideal strength later using PixelMath, unless you know the exact amount of noise you want to remove in advance.

Model Version

• v1 (RGB Only): Works only on RGB images.
• v2 (RGB and Grayscale): Supports both RGB and monochrome images and is generally recommended for most cases.

Both models are trained primarily on monochrome (CCD) sensor data.

🚨 For color sensor data (Bayer matrix), use CFA Drizzle integration before applying DeepSNR.

Linear Data (Checkbox)

If checked, DeepSNR automatically applies a stretch to the image before denoising and de-stretches it afterward.

If no STF (Screen Transfer Function) is applied by user, it will apply auto-stretch the image before processing. This usually works well, but results may vary, so better set the STF manually.

Usage Tips

• ✅ Apply only to high-quality raw data immediately after integration.
• ✅ Works best on uncorrelated noise. If your data has structured noise (e.g., dithering issues), artifacts may appear.
• ✅ Tolerates hot and cold pixels fairly well.
• ✅ v2 handles column artifacts better, though it’s recommended to correct these before denoising.
• 🚫 Do not use on heavily processed images—this tool is not meant for aggressive noise reduction after post-processing.

General Notes on Usage

• The model is trained on non-linear (stretched) images. If the linear flag is checked, the image will be stretched before processing and de-stretched afterward, which may lead to undesired effects such as clipping in the darkest regions.
• When working with linear images, ensure that an appropriate Screen Transfer Function (STF) is applied before denoising. DeepSNR will use auto-stretch if the linear flag is checked.
• The stretch level is important—the model processes the image exactly as displayed, so extreme highlights or shadows may not be efficiently processed. Ideally, you want the image you denoise to be as close to the final result as possible.
• If your image has a high noise level, there is a risk of "hallucinated" stars. In such cases, consider using lighter stretch.

Known Problems

• ⚠ When applied to linear images, high-intensity regions (such as stars) might be slightly damaged. Better results are usually achieved on stretched images. Be aware of this effect when processing data and preferrably check out the difference between the original and denoised images if you really care about the quality of your data.