摘要
针对现有的图像分割技术在小样本量数据集上容易过拟合,不能有效分割缺损图像的问题,提出了一种自约束图像分割方法.首先,基于传统无监督水平集图像分割方法,提出一种可微分水平集层.第二,将可微分水平集层嵌入U-Net等有监督图像分割模型中,使得水平集方法对函数的拓扑约束,可以随着梯度反向传播过程,对卷积参数起到约束作用.实验结果表明,在MNIST和Fashion-MNIST简单数据集上,本文方法的分割准确率比CV等基于水平集的方法分别提升8.3%和11.7%,比U-Net等分割网络分别提升7.5%和15.6%;在背景复杂的Weizmann horse数据集上准确率较基于水平集的方法提高54.9%,较U-Net等分割网络提升13.4%,显示出本文方法在小样本缺损图像数据集上的有效性与鲁棒性.
Aiming at the problem that the existing image segmentation technology is easy to overfit on the small sample size data set and cannot effectively segment the defective image,a self-constrained image segmentation method is proposed.First,based on the traditional unsupervised level set image segmentation method,a differentiable level set layer is proposed.Second,the differentiable level set layer is embedded in the supervised image segmentation model such as U-Net,so that the level set method can restrict the topological constraints of the function and constrain the convolution parameters along with the gradient back propagation process.Experimental results show that,on MNIST and Fashion-MNIST simple data sets,the segmentation accuracy of the proposed method is 8.3%and 11.7%higher than that of the level set based method such as CV,and 7.5%and 15.6%higher than that of U-Net and other segmentation networks,respectively;Compared with the level set-based method,the accuracy rate on the Weizmann horse data set with complex background is improved by 54.9%,and it is improved by 13.4%compared with the segmentation network such as U-Net.Shows the effectiveness and robustness of this method on small sample defect image data sets.
作者
刘剑超
相洁
张玲
李钢
LIU Jian-chao;XIANG Jie;ZHANG Ling;LI Gang(College of Computer Science&Technology,Taiyuan University of Technology,Taiyuan 030024,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2022年第8期1732-1738,共7页
Journal of Chinese Computer Systems
基金
山西省自然科学基金项目(201901D111091)资助
山西省高校科技创新项目(JYTKJCX201943)资助.
关键词
图像分割
水平集
深度学习
神经网络
污损图像
image segmentation
level set
deep learning
neural network
fouling image