摘要
针对空间科学实验中复杂场景下细胞图像难以精细准确分割的问题,提出了基于Mask R-CNN的实例分割新算法——基于密集特征金字塔的实例分割网络(DFP-Mask)。该算法在特征金字塔网络自顶向下的特征传输过程中以密集连接的方式控制多尺度特征图间的信息传递,将高层语义结构信息传递至所有低层特征,提高低层特征的语义理解能力,从而提升多尺度特征的目标识别能力。实验选用天舟一号小鼠肝卵圆细胞图像数据,数据集中包涵200张背景复杂且有实例交叠的图片。实验结果表明:与Mask R-CNN相比,DFP-Mask在多个评价指标和视觉分割效果上表现更优,其中准确率提高了2.03%,召回率提高了3.77%,平均精确率mAP提高了1%。DFP-Mask可应用于更多空间科学实验对象的数量、形态、生长过程等表型特征的提取。
It is difficult to finely and accurately segment cell images under complex scenes in space science experiments.To solve this problem,a new instance segmentation algorithm based on Mask R-CNN network and Dense Feature Pyramid(DFP-Mask)was proposed.In the top-down feature transmission process of the Feature Pyramid Networks(FPN),the algorithm controlled the information transfer between multi-scale feature maps in a densely connected manner,transferred high-level semantic structure information to all low-level features,and improved the semantic understanding ability of low-level features,thereby improved the target recognition ability of multi-scale features.The Tianzhou-1 mouse liver oval cell data set containing 200 pictures with complex backgrounds and overlapping examples was used in the study.The results showed that compared with Mask R-CNN,DFP-Mask performed better in multiple evaluation indicators and visual segmentation effects.The Precision increased by 2.03%,the Recall increased by 3.77%,and mAP increased by 1%respectively.DFP-Mask can be applied to the extraction of phenotypic features such as the quantity,shape,and growth process of more space science experiment objects.
作者
董高君
许乐乐
马忠松
于歌
DONG Gaojun;XU Lele;MA Zhongsong;YU Ge(Technology and Engineering Center for Space Utilization,Key Laboratory of Space Utilization,Chinese Academy of Sciences,Beijing 100049,China;University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《载人航天》
CSCD
北大核心
2021年第2期169-174,共6页
Manned Spaceflight
基金
载人航天领域预先研究项目(18051030301)。
关键词
深度学习
实例分割网络
细胞分割
多尺度特征
密集特征金字塔
细胞图像
deep learning
instance segmentation network
cell segmentation
multi-scale features
Dense Feature Pyramid
cell image