摘要
为了提高语义分割的精确性和完整性,提出一种单阶段弱监督学习的语义分割方法。首先,在像素信息保持神经网络的基础上,提出了归一化全局加权池化,并将焦点遮罩惩罚引入分类得分函数中,以促进图像语义的完整性。然后,提出像素自适应遮罩精细化,对外观线索相关的粗糙遮罩预测进行修正,更新后的遮罩作为伪真实数据进行分割。并在此基础上,利用随机门将深层特征与浅层特征相结合。实验在格萨尔唐卡数据集和Pascal VOC 2012数据集上进行,实验结果表明,与一些优秀方法相比,所提方法具有较高的交并比(IoU),且消融实验也表明各模块可以明显提高语义分割性能。
In order to improve the accuracy and integrity of semantic segmentation,a single-stage learning with weak supervision method is proposed.Firstly,based on the IRNet,the normalized global weighted pooling technique is proposed,and a focus mask penalty is introduced into the classification score function to promote the semantic integrity of the image.Then,a pixel adaptive mask refinement strategy is proposed to modify the rough mask predictions related to appearance cues,and the updated masks are used as pseudo-real data for segmentation.On this basis,the random gate is used to merge the deep features with the shallow ones.Experiments are carried out on Gesar Thangka data set and Pascal VOC 2012 data set.And the experimental results show that compared with state-of-art methods,the proposed method obtains higher intersection over union(IoU).The ablation experiments also confirm that each module in the proposed model can significantly improve the semantic segmentation performance.
作者
肖衡
Xiao Heng(School of Information&Intelligence Engineering,University of Sanya,Sanya 572022,China;Academician Workstation of Chen Guoliang,Sanya 572022,China)
出处
《国外电子测量技术》
北大核心
2021年第12期30-36,共7页
Foreign Electronic Measurement Technology
基金
海南省自然科学基金(619QN243)
海南省自然科学基金高层次人才项目(2019RC257)资助
关键词
语义分割
自适应精细化遮罩
分类得分
交并比
随机门
semantic segmentation
adaptive refinement mask
classification score
intersection over union
random gate
作者简介
肖衡,硕士,副教授,主要研究方向为机器学习、数字信号处理等。E-mail:melody30099@163.com
