| The detection and study of astronomical transients,which carry a wealth of information about the celestial bodies and the evolution of the universe,is of great scientific value.Most of the peak radiation from astronomical transients is in the highenergy band,making space-based telescopes more suitable for the detection of astronomical transients than ground-based telescopes due to the influence of the Earth’s atmosphere.Constrained by satellite-ground communication links,the engineering cost of transmitting large amounts of images taken by space-based telescopes back to the ground is high,and a significant proportion of the data is invalid.Constrained by the performance of on-board computers,it is difficult to implement space-based Convolutional Neural Network(CNN)transients detection algorithms that rely on the powerful computing power on the ground.Therefore,the construction of a lightweight transients detection model based on CNN and its deployment in an on-board finite computing platform are of great significance for future space applications of transients detection.The main research work and achievements covered in this paper include:(1)This paper investigates the current research on the techniques of transients detection and analyses the advantages and disadvantages of different transients detection techniques and the main types of targets to be detected.This paper searches for a reliable data set of astronomical transients with large data sample sizes,and performs relevant pre-processing operations on the data set so that it can be used as input data for the CNN model for training and optimization.(2)In this paper,various deep Convolutional Neural Networks such as VGG16,Inception V3,Res Net32 and Mobile Net V3 are constructed based on the characteristics of the astronomical transients images in the pre-processed data set,and the performance of the models is continuously optimized and improved after the initialization of the relevant parameters.The experimental results show that the Inception V3 and Mobile Net V3 models have good detection results on the astronomical transients data set,with 94.35% and 90.22% accuracy respectively,while the VGG16 and Res Net32 models have less than 90% accuracy.From the perspective of model hardware deployment,the four models mentioned above are relatively large in terms of number of parameters and computational effort,and the models are complex,which make them not suitable for deployment on an on-board finite computing platform.(3)This paper analyzes the structure of the above four Convolutional Neural Network models and their corresponding performance,and constructs three lightweight CNN-based transient source detection methods,including Shallow-level CNN Transient Detection(SCTD),Group Shallow-level CNN Transient Detection(GSCTD)and Depthwise separable Shallow-level CNN Transient Detection(DSCTD).It has been experimentally verified that,compared with the Deep-Hits model,which currently has a relatively high accuracy of transients detection,the DSCTD model constructed in this paper is less than 1/5 of the Deep-Hi TS model in terms of number of parameters and less than 1/32 of the Deep-Hi TS model in terms of computation,and achieves an accuracy of 98.44%,which is only 1.01% lower than that of the Deep-Hi TS model,and is therefore more suitable for deployment to a satellite-based finite computing platform to achieve transients detection at terminals.(4)This paper completes the construction of the hardware platform environment,optimizes the DSCTD model and deploys it to the embedded ARM hardware platform via the Tengine framework,and verifies the performance of the deployed model for transients detection.The experimental validation shows that the accuracy of the deployed model’s transients detection reaches 98.43% and the running time of the input500 astronomical transients images is 7.77 s,which can better meet the requirements of the transients detection task in terms of accuracy and real-time performance. |
PDF Full Text Request