摘要： 在进行超新星目标检测时，图像背景复杂、目标较小以及正负样本不平衡导致图像对比不明显和特征提取难度大等问题。为此，从数据合成、特征提取网络优化等方面对Faster R-CNN算法进行改进，提出一种超新星目标检测方法。将每组图像进行合成以提高图像的对比度。针对特征提取难度大的问题，使用深度残差网络提取合成图像的特征，并将顶层特征依次与低层特征相融合，构建特征金字塔网络，使每一层网络都具有较强的语义信息。采用在线难例挖掘方法对高损失样本进行训练，以处理正负样本不平衡的问题，从而提高算法的检测性能。实验结果表明，与原始Faster R-CNN算法相比，该算法的Score与F1值分别提高8.51%和45.52%，且其检测性能与泛化能力均较高。

关键词: 超新星, 神经网络, 目标检测, 特征金字塔网络, 在线难例挖掘

Abstract: Existing supernova detection methods suffer from poor contrast of images and difficult feature extraction caused by the complex image background, small size of object,and imbalance between positive and negative samples.To address the problem,this paper proposes a supernova detection method by improving the Faster R-CNN algorithm from the perspective of data integration and optimization of feature extraction network.The method synthesizes each group of images to improve their contrast.To reduce the difficulty of feature extraction,the deep residual network is used to extract the features of the synthesized images,and the top layer features are fused with the lower layer features to construct the feature pyramid network,so that each layer of the network has strong semantic information.At the same time,the Online Hard Example Mining(OHEM) method is used to train the high loss samples to deal with the imbalance between the positive and negative samples,so as to significantly improve the detection performance of the algorithm.Experimental results show that compared with the original Faster R-CNN algorithm,the proposed algorithm has improved the Score by 8.51% and F1 score by 45.52%.It has better detection performance and generalization ability.

Key words: supernova, neural network, object detection, feature pyramid network, Online Hard Example Mining(OHEM)

中图分类号: 

• TP391