Arp 204 · DeepCR Demo
Drag the handle to compare the raw frame and the DeepCR output. Notice how the deep learning model effectively removes cosmic rays while preserving the galaxy structure.
- Deep Learning detection trained on massive datasets.
- Native 32-bit support without rescaling issues.
- Simple Parameters: just choose a model and threshold.
Raw Input
Cleaned
PixInsight Interface
A streamlined interface for easy operation. Select your model, set the threshold, and execute.
Recommended Settings
DeepCR offers specific models and thresholds for different data types.
- 32-bit Float (Common): Optimal Preset (WFC3-UVIS, Threshold 0.1).
- Space Telescope (HST/JWST): Optimal or ACS Default (Threshold 0.1-0.2).
- Ground-Based Long Exp: Optimal (Threshold 0.10-0.15).
- Faint Sources: Conservative (Threshold 0.2).
Comparison with LACosmic
While LACosmic is excellent, DeepCR leverages deep learning for superior results in many cases.
- Method: Deep Learning vs Edge Detection.
- Accuracy: Higher accuracy with fewer false positives on faint stars.
- Speed: Comparable (~10-15 sec).
- Simplicity: 2 simple parameters vs 8+ complex ones.
Scientific Background
This module is based on the work of Zhang & Bloom (2020). It uses a U-Net architecture trained on over 15,000 labeled HST images to identify cosmic rays with high precision.
Citation: Zhang, K., & Bloom, J. S. (2020). Identifying Cosmic Rays in Astronomical Images Using Deep Learning. The Astrophysical Journal, 889(1), 24.