A Supernova Detection Implementation based on Faster R-CNN
Tianyuan Wu · 2020
Object detection is the basis of many computer vision applications. It combines two tasks which are object classification and location. With the development of astronomical observation technology, using object detection methods to find more novae and supernova is becoming an interesting and practical issue. In this paper, a model-based supernova object detection framework including pseudo color image compositing technology is proposed by using faster R-CNN. And by balancing the uneven positive and negative sample, data enhancement, feature extraction, focal loss modification and neural network training, the model gives a suitable method for supernova detection. The deep learning network is realized by Tensorflow and the network test results show that the proposed method can achieve better accuracy under the given prediction standard calculation method. The training and testing sample are provided by the Popular Supernova Project(PSP).