摘要
针对获得训练数据集代价高昂问题,提出了一种用于图像显著性检测的弱监督新方法,在训练网络模型时仅使用图像级标签。方法分为两个阶段,在第一阶段,根据图像级标签训练分类模型,获得前景推断图;在第二阶段,对原图像进行超像素块处理,并与阶段一得到的前景推断图进行融合,从而细化显著对象边界。算法使用了现有的大型训练集和图像级标签,未使用像素级标签,从而减少了注释的工作量。在四个公共基准数据集上的实验结果表明,性能明显优于无监督的模型,与全监督模型相比也具有一定的优越性。
Aiming at the high cost of obtaining the training data set,this paper proposed a new weak supervision method for image saliency detection.It only used the picture-level label when training the network model.It divided the method into two stages.In the first stage,it trained the classification model according to the picture-level label to obtain the foreground inference graph.In the second stage,it processed the original image by super-pixel block and merged with the foreground inference graph obtained in phase one,thus refined significant object boundaries.The algorithm used existing large training sets and image-level tags,eliminated the use of pixel-level tags,which reduced the amount of annotation work.The experimental results on the four common benchmark datasets show that the performance is significantly better than the unsupervised model,and it has certain advantages compared with the full-supervised model.
作者
谭台哲
轩康西
曾群生
Tan Taizhe;Xuan Kangxi;Zeng Qunsheng(College of Computer,Guangdong University of Technology,Guangzhou 510006,China;Heyuan Guanggong Collaborative Innovation Research Institute,Heyuan Guangdong 517000,China)
出处
《计算机应用研究》
CSCD
北大核心
2020年第2期601-605,共5页
Application Research of Computers
关键词
深度学习
弱监督
显著性检测
超像素
deep learning
weak supervision
significance detection
superpixel
作者简介
谭台哲(1970-),男,山东莱阳人,副教授,博士,主要研究方向为机器学习与大数据处理、图像处理与计算机视觉;通信作者:轩康西(1993-),男,硕士,主要研究方向为深度学习、图像处理(1056808552@qq.com);曾群生(1993-),男,硕士,主要研究方向为深度学习、图像处理.
