摘要
本文在DeepLab-V2模型的基础上提出了一种扩张卷积优化的图像语义分割模型:DODNet.在DODNet中,本文采用多混洗块扩张卷积(MSDC)替换ResNet-101中的扩张卷积,通过通道混合(ChannelShuffle)为每个特征点添加相邻特征点的部分通道(channel)信息,缓解了连续级联扩张卷积带来的网格化效应;同时,本文采用Vortex-Conv模块替换掉空洞空间金字塔池化(ASPP),以增加扩张卷积对感受野信息的利用,通过像素点周围局部信息的有效融合,获取了更精确的分割结果.本文在PASCAL VOC12数据集上对算法模型进行了训练和验证.实验结果表明,DODNet模型相比DeepLab-V2模型获得了4.23%的分割精度提升,同时模型参数量减少了11.2M,计算量降低了12.2B.并且与DeepLab-V3相比,DODNet模型也获得了0.32%的分割精度提升,以及7.1M的参数量和7.7B计算量的下降.
In this paper,we propose a novel image semantic segmentation model with optimized dilated convolution,named DODNet(DeepLab with optimized dilated convolutions).We build DODNet based on DeepLab-V2.In DODNet,we design a new Multi Shuffle-Block Dilated Convolution(MSDC)module to replace the original dilated convolution in ResNet-101,in which a Channel-Shuffle block is applied to enhance the use of partial channel information of adjacent feature points to alleviate the gridding effect caused by continuous cascade dilated convolution.In order to have a more effective utilization of contextual information in receptive field,another new designed Vortex-Conv module is brought in to replace the original use of Atrous Spatial Pyramid Pooling(ASPP)module in DeepLab-V2,which helps to achieve more accurate segmentation results through an enhanced fusion of local and global information.We train and test our model on PASCAL VOC12 dataset.Our experimental results show that,compared to DeepLab-V2,the proposed DODNet model has obtained a 4.23%outperformance in segmentation accuracy while at the same time having its parameters and calculations been reduced by 7.01 M and 7.6 Brespectively.Compared with DeepLab-V3,the proposed DODNet model can also have a 0.32%performance gain,as well as 7.1 Mand 7.7 Breduction in the amount of parameters and calculations.
作者
祖朋达
李晓敏
陈更生
许薇
ZU Penda;LI Xiaomin;CHEN Gengsheng;XU Wei(State Key Laboratory of ASIC&System,Fudan University,Shanghai 201203,China)
出处
《复旦学报（自然科学版）》
CAS
CSCD
北大核心
2020年第5期585-598,607,共15页
Journal of Fudan University：Natural Science
