摘要
恒星星系的精准识别是开展很多天文海量数据分析和处理任务的基础。受环境因素影响,采集到的暗星体观测数据使得恒星和星系差异不明显,而且暗星体数据量较小,给分类带来困难,所以在恒星/星系识别任务中准确地识别暗星体成为近年来研究的重点。提出一种用于极暗星体识别的硬调整孪生网络模型,解决了小样本问题和困难样本挖掘问题,将极暗星体的识别效果较目前最好结果提升了8百分点左右;同时提出用于暗星体和亮星体识别的代价敏感模型,解决了数据量充足条件下的困难样本挖掘问题,暗星体和亮星体的识别效果较目前最好结果分别提升了1百分点和0.1百分点。
The precise identification of the stars and galaxy is a prerequisite for many other astronomical tasks.Due to environment,the difference between faint stars and dark galaxies is not obvious,and the amount of faint stars and faint galaxies is small.The classification of the faint samples is difficult.Therefore,identifying faint stars and galaxies accurately has become the focus in recent researches.In this paper,we propose the Siamese network with hard manual controlling for the recognition of extreme faint stars and galaxies,which is suitable for few-shot learning and hard example mining.Comparing the performance of our model with the state of art,we find that the completeness of extreme faint examples increased by 8 percentage points.We also propose a cost sensitive model for the recognition of faint or bright stars and galaxies,which is opportune for hard example mining under sufficient data.We find that the completeness of faint and bright examples increased by 1 percentage points and 0.1 percentage points.
作者
张士川
郑小盈
Zhang Shichuan;Zheng Xiaoying(School of Microelectronics,University of Chinese Academy of Sciences,Beijing 100049,China;Shanghai Advanced Research Institute,Chinese Academy of Sciences,Shanghai 201210,China)
出处
《计算机应用与软件》
北大核心
2021年第12期149-154,186,共7页
Computer Applications and Software
基金
国家自然科学基金项目(U1831118)
上海市自然科学基金项目(19ZR1463900)。
关键词
恒星/星系分类
小样本
孪生网络
困难样本挖掘
代价敏感
Star/Galaxy classification
Few-shot learning
Siamese network
Hard examples mining
Cost sensitive
