# Effective Image Differencing with ConvNets for Real-time Transient Hunting

**Authors:** Nima Sedaghat, Ashish Mahabal  
**arXiv:** [1710.01422](https://arxiv.org/abs/1710.01422)  
**Date:** 2017-10-04  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1710.01422v1/

## Key points

- Large sky surveys increasingly rely on image subtraction pipelines for real-time transient detection
- Challenges include dealing with varying Point Spread Functions, brightness fluctuations, artifacts from saturated stars and matching errors
- A novel deep-learning approach integrates essential steps of traditional pipeline into a single real-time convolutional network
- Trained approach demonstrates remarkable speed and efficiency in performing multiple tasks simultaneously
- Advantages particularly pronounced for multi-CCD and fast surveys like ZTF and LSST
- Deep learning technology streamlines process and enhances accuracy of transient detection
- Refined methodology opens up new possibilities for uncovering rare or elusive astronomical phenomena

## Summary

, , , , 

In the realm of large sky surveys, the reliance on image subtraction pipelines for real-time transient detection is becoming increasingly prevalent. However, this process comes with its own set of challenges, including dealing with varying Point Spread Functions (PSF), minor brightness fluctuations in numerous sources, and artifacts stemming from saturated stars and matching errors. Moreover, the use of a reference image that is deeper than individual images can introduce differences in noise characteristics that may lead to additional artifacts. To address these issues, a novel deep-learning approach to transient detection has been introduced. This innovative method integrates all the essential steps of a traditional image subtraction pipeline – such as image registration, background subtraction, noise elimination, PSF matching, and subtraction – into a single real-time convolutional network. Once trained, this approach demonstrates remarkable speed and efficiency in performing multiple tasks simultaneously. The advantages of this advanced technique are particularly pronounced for multi-CCD and fast surveys like ZTF and LSST. By streamlining the process and enhancing the accuracy of transient detection through deep learning technology, researchers are able to expedite the identification and analysis of celestial events in large-scale sky surveys. The potential impact of this refined methodology extends beyond just improving efficiency; it also opens up new possibilities for uncovering rare or elusive astronomical phenomena that may have previously gone undetected.