基于深度卷积神经网络的前景对象图像分割模型FOSegNet  

FOSegNet:foreground object segmentation model using deep convolution neural networks

作　　者：吴彬 杨戈 陈海洋 WU Bin;YANG Ge;CHEN Hai-yang(Branch School of Zhuhai,Beijing Normal University,Zhuhai 519000,China;Gradute School of Beijing Normal University,Beijing 100875,China)

机构地区：[1]北京师范大学珠海分校,广东珠海519000 [2]北京师范大学研究生院,北京100875

出　　处：《东北师大学报（自然科学版）》2020年第4期74-81,共8页Journal of Northeast Normal University(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61272364);广东高校省级重大科研项目(201612008QX,2016KTSCX167,2017KTSCX207);广东大学生科技创新培育专项项目(pdjh2019b0581).

摘　　要：提出一个端到端的基于深度卷积神经网络DCNNs(Deep Convolutional Neural Networks)的监督学习模型FOSegNet(Foreground Object Segmentation Networks),用于逐像素的前景对象图像分割.首先能有效地扩大滤波器视野的扩张卷积代替常用卷积,以便加入更多的上下文信息而不增加参数数量;然后提出并应用分流聚合模块SFM(Shunt-Fuse Module),在多尺度上鲁棒地分割物体,增强分割模型的泛化能力;最后级联DCNNs、分流聚合模块和概率图模型作为分割模型的特征提取器,实现模型端到端的训练与分割.实验结果表明,FOSegNet模型在MIT Object Discovery和ImageNet-Segmentation数据集上均超过了众多前景对象分割模型的性能表现,在PASCAL VOC 2012数据集上的分割表现也优于众多语义分割模型.An end-to-end supervised learning model FOSegNet(Foreground Object Segmentation)based on Deep Convolutional Neural Networks(DCNNs)for pixel-by-pixel foreground object image segmentation is proposed.Firstly,dilated convolution is used instead of common convolution,because it can effectively expand the receptive filed of the filter,and add more context information without increasing the number of parameters.Then,Shunt-Fuse Module(SFM)is proposed and applied to robustly segment objects on multi-scale to enhance the generalization ability.Finally,DCNNs、SFM and graph model are cascaded and used as feature extractors of segmentation model to achieve end-toend training and segmentation of the model.The experimental results show that our model surpasses the performance of many advanced foreground object segmentation models on both MIT Object Discovery and ImageNet-Segmentation datasets.And our model also outperforms many semantic segmentation models on PASCAL VOC 2012 dataset.

关 键 词：卷积神经网络 分流聚合模块 前景对象分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]