Application And Research Of Vision Algorithm In Pulsar Candidate Recognition

So-called pulsars are relatively rare neutron stars that send out radio radiation that sweeps across Earth and is detected by telescopes.Pulsars are of great scientific significance as detectors of space-time and the state of interstellar medium and matter.In 2016,a 500-meter-aperture spherical radio telescope was established in Guizhou,with a sensitivity more than 2.5 times that of the world's second largest telescope,and about 500 pulsars were discovered in two years.Exceeds the 15-year survey results of the Arecibo Telescope in the United States.However,compared to the number of pulsars estimated by scientists,500 is just a drop in the ocean,so the search for pulsars still requires faster and more accurate algorithms to continuously optimize the search process.Today,with the rapid advancement of algorithms in the field of vision,many astronomers apply image technology to solve problems.In previous work,many astronomers have used deep learning for candidate screening,mainly using convolutional neural networks to extract features,or using generative adversarial networks to generate samples.This paper is similar to the latter work,mainly focusing on pulsars The imbalance of the dataset,the starting point is the expansion of the dataset and the design of the loss function,focusing on the design of the enhancement method for the pulsar dataset and the design of the loss function.Second,unlike the previous semi-supervised methods,this paper attempts to generate images of pulsars using only label data,and achieves certain results.The following are the main work and innovations accomplished by the paper:1.First of all,this paper introduces the process of pulsar candidate identification,the problems encountered in pulsar candidate screening,and the characteristics of pulsar data sets.In response to these problems,the development of pulsar image recognition methods in recent years has been sorted out.By reading the literature on related algorithms in recent years,we can learn from the innovative methods in the current image recognition field.The first work focuses on two types of methods,namely algorithms The principles and methods of cost-sensitive learning and the basic enhancement techniques are discussed respectively at the level and sample level,and a very important part of the subsequent application is described,that is,the method of mining difficult samples,and the focus loss is introduced.In the following research,starting from sample enhancement,the problems caused by the introduction of image enhancement in pulsar candidate image recognition are solved by introducing difficult sample mining and cost-sensitive assistance.2.Another idea to solve the problem of unbalanced samples is to learn the distribution of data,and then sample the distribution to supplement the training set.This paper studies the related algorithms of generative adversarial networks in recent years,introduces the historical development of related algorithms,and analyzes the characteristics of the algorithms.And combined with the pulsar data set to improve the algorithm.In the fourth chapter,these methods are tested and the results are analyzed.At the end of the paper,the paper reviews the full text,analyzes the experimental data,what results are obtained,and further analyzes the experimental results,makes assumptions about the problems that may still exist,and looks forward to the follow-up research.