An Intelligent Detection Method of Astronomical Transients Based on Lightweight CNN Model
-
摘要: 天文暂现源携带了关于天体本质及演化过程的丰富信息,对暂现源进行探测与研究具有极为重要的科学价值。天文暂现源的辐射峰值大多在X射线或伽马射线,天基望远镜对这些高能波段的观测优势是地基望远镜无法比拟的,更适合于暂现源观测。但由于星载计算机的性能约束,很难实现依托于地面强大算力的复杂检测算法。针对以上问题,提出了基于轻量化卷积神经网络(CNN)模型的天基暂现源检测算法,并在嵌入式ARM平台上实现了模型部署。实验结果表明,本文提出的轻量化CNN暂现源检测算法的模型复杂度和计算量不及Deep Hits算法的1/4,准确率达到96.52%,可应用于星载有限算力平台,实现未来的天基暂现源实时检测。Abstract: Astronomical Transients carry rich information about the nature and evolution of celestial bodies, and their detection and research have extremely important scientific value. Most of the radiation peaks of astronomical transients are in X-rays or Gamma rays. The observation advantages of space-based telescopes in these high-energy bands are unmatched by ground-based telescopes, and they are more suitable for transients observation, but due to the constraints of the performance of on-board computers, it is difficult to implement complex detection algorithms that rely on the powerful ground computing power. In response to the above problems, a transient detection algorithm is proposed based on the lightweight Convolutional Neural Network (CNN) model, and the model deployment is implemented on the embedded ARM platform. The experimental results show that the model complexity and computational complexity of the lightweight CNN transients detection algorithm proposed are less than 1/4 of the Deep Hits algorithm, while the accuracy rate can reach 96.52%, and it can be applied to a space-borne limited computing power platform to realize real-time detection of space-based transients in the future.
-
图 2 预处理后的图像数据。每一组图像从左到右依次为模板图像、科学图像、差分图像、信噪比图像。label=1.0代表真实暂现源,label=0.0代表伪暂现源
Figure 2. Preprocessed image data. Each group of images from left to right are template image, scientific image, differential image, signal-to-noise ratio image. label=1.0 represents the real transients, label=0.0 represents the pseudo transients
图 3 轻量化CNN网络模型架构(C代表卷积层、P代表池化层、FC代表全连接层,池化层采用最大池化策略,激活函数采用RELU函数)
Figure 3. Lightweight CNN network model architecture (C represents the convolutional layer, P represents the pooling layer, FC represents the fully connected layer, the pooling layer adopts the maximum pooling strategy, and the activation function adopts the RELU function)
表 1 预测结果混淆矩阵
Table 1. Confusion matrix of prediction results
实际情况 预测结果 正例 反例 正例 6901(PT) 299(NF) 反例 151(PF) 6956(NT) -
[1] YU Hai. Applications of Transients and Other Observations in Cosmology[D]. Nanjing: Nanjing University, 2019 [2] ZHAO Yifei. Astronomical Transient Source Recognition Based on Deep Learning and Raspberry Pi[D]. Taiyuan: Taiyuan University of Technology, 2019 [3] PENG Bin, LI Zheng, LIU Yong, et al. Automatic code comments generation method based on convolutional neural network[J]. Computer Science, 2021, 48(12): 117-124 doi: 10.11896/jsjkx.201100090 [4] CABRERA-VIVES G, REYES I, FÖRSTER F, et al. Deep-HiTS: rotation invariant convolutional neural network for transient detection[J]. The Astrophysical Journal, 2017, 836(1): 97 doi: 10.3847/1538-4357/836/1/97 [5] SEDAGHAT N, MAHABAL A. Effective image differencing with convolutional neural networks for real-time transient hunting[J]. Monthly Notices of the Royal Astronomical Society, 2018, 476(4): 5365-5376 doi: 10.1093/mnras/sty613 [6] CONNOR L, VAN LEEUWEN J. Applying deep learning to fast radio burst classification[J]. The Astronomical Journal, 2018, 156(6): 256 doi: 10.3847/1538-3881/aae649 [7] HUANG Tianjun. Research on Detecting Transients and Variable Sources in AST3-2 Survey[D]. Hefei: University of Science and Technology of China, 2019 [8] CHEN Xiaotong. Research on aviation sheet metal parts missing detection technology based on VGG technology[J]. Modern Industrial Economy and Informationization, 2021, 11(9): 219-220 [9] XU L Y, GAJIC Z. Improved network for face recognition based on feature super resolution method[J]. International Journal of Automation and Computing, 2021, 18(6): 915-925 doi: 10.1007/s11633-021-1309-9 [10] RAN H H, WEN S P, SHI K B, et al. Stable and compact design of memristive GoogLeNet neural network[J]. Neurocomputing, 2021, 441: 52-63 doi: 10.1016/j.neucom.2021.01.122 [11] TAN Y H, JING X J. Cooperative spectrum sensing based on convolutional neural networks[J]. Applied Sciences, 2021, 11(10): 4440 doi: 10.3390/app11104440