Development of transient detection method and GPU-accelerated image reduction pipeline
Development of transient detection method and GPU-accelerated image reduction pipeline
复制标题
瞬态检测方法和 GPU 加速图像缩减管道的开发
DOI：
10.1117/12.2561258
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
Hosokawa Ryohei
中科院分区：
文献类型：
--
作者：
Murata Katsuhiro;Iida Kota;Niwano Masafumi;Yatsu Yoichi;Kawai Nobuyuki;Oeda Motoki;Shiraishi Kazuki;Adachi Ryo;Ogawa Futa;Toma Sayaka;Hosokawa Ryohei
引力波光学对应物的识别是天文学中最令人兴奋的课题之一。由于LIGO/Virgo的典型天图误差区域远大于光学望远镜的视场，因此通过光学望远镜的后续观测来搜索和快速识别光学对应物非常重要。如何在大量的观测图像中快速准确地检测出瞬态信号是一个重要的问题。 受此启发，我们正在开发使用卷积神经网络（CNN）的瞬态检测方法。我们构建了基于CNN的分类器，旨在分离瞬态图像，包括瞬态源的图像，和非瞬态图像。输入数据是一对观察图像和参考图像。这里我们采用MITSUME 50 cm望远镜拍摄的图像作为观测图像，Pan-STARRS图像作为参考图像。我们用200 Mpc内77个背景星系的10，000多张图像训练它。通过在不同位置和亮度的观测图像中加入人工点源，得到人工瞬变图像的训练数据。我们用测试数据对分类器的性能进行了测试，发现分类准确率在90%以上。 此外，我们正在开发一个具有GPU（图形处理单元）的高速图像缩减流水线，用于实时分析观察到的图像。为了加速图像缩小，管道使用CuPy（用于GPU上数值计算的Python库）并最小化fits I/O。我们发现，对于240组1024 x 1024像素的图像，流水线的缩减速度比IRAF快30倍。 在这次演讲中，我们将介绍瞬态检测方法和GPU加速的图像缩减流水线的发展现状。我们还将介绍我们将它们安装到位于Akeno和冈山的MITSUME 50 cm望远镜系统中的计划，这些望远镜已经对伽马射线爆发、引力波和X射线双星进行了光学跟踪观测。
Identification of optical counterparts of gravitational waves (GWs) is one of the most exciting topics in astronomy. Since a typical sky map error region of the LIGO/Virgo is much larger than the field-of-view of optical telescopes, it is important to search and rapidly identify optical counterparts through follow-up observation by optical telescopes. The method of rapid and accurate transient detection in huge set of observed images is important. Motivated by this, we are developing transient detection method with convolutional neural network (CNN). We constructed CNN-based classifier designed to separate a transient image, an image including a transient source, and non-transient image. The input data is a pair of an observed image and a reference image. Here we adopt an image taken by MITSuME 50 cm telescope as observed and Pan-STARRS image as reference. We trained it with more than 10,000 images of 77 background galaxies within 200 Mpc. The training data with artificially transient images is made by adding an artificial point source into an observed image with various positions and luminosities. We tested the performance of the classifier with test data and found that the classification accuracy is more than 90%. Furthermore, we are developing a high-speed image reduction pipeline with GPU (Graphics Processing Unit) for real-time analysis of observed images. To accelerate image reduction, the pipeline uses CuPy (a python library for numerical calculation on the GPU) and minimize fits I/O. We found that the reduction speed of the pipeline achieves 30 times faster than IRAF for 240 set of 1024 x 1024 pixel images. In this talk, we will introduce the current status of the development of the transient detection method and the GPU-accelerated image reduction pipeline. We will also introduce our plan of installation of them into the systems of MITSuME 50cm telescopes in Akeno and Okayama which have performed optical follow-up observations of gamma-ray bursts, gravitational waves and X-ray binaries.