基于改进DeepLabv3的自然图像语义分割算法  

Image semantic segmentation algorithm based on improved DeepLabv3

作　　者：赵晓[1] 王若男 杨晨[1] 李玥辰 ZHAO Xiao;WANG Ruo-nan;YANG Chen;LI Yue-chen(School of Electronic Information and Artificial Intelligence,Shaanxi University of Science&Technology,Xi′an 710021,China)

机构地区：[1]陕西科技大学电子信息与人工智能学院,陕西西安710021

出　　处：《陕西科技大学学报》2024年第2期182-188,共7页Journal of Shaanxi University of Science & Technology

基　　金：国家自然科学基金项目(61971272,61601271);陕西科技大学博士科研启动基金项目(2019BJ-27)。

摘　　要：针对DeepLabv3模型对自然图像语义分割时存在的图像局部细节信息丢失导致的误分割和物体边缘分割不完整的问题,提出了一种改进DeepLabv3模型的自然图像语义分割网络,能够以更高的准确率实现自然图像的语义分割.首先,使用ResNet101作为骨干网络进行特征提取,把ResNet101网络最后两层提取到的特征图输入到设计的ACMix多重融合模块(ACMix Multiple Fusion Module,AMFM)中,有效获取不同尺度的空间特征信息,将融合之后的结果作为空洞空间金字塔池化模块(Atrous Spatial Pyramid Pooling,ASPP)的输入.其次,添加辅助分支模块(Auxiliary Branch Module,ABM),将ResNet101网络第三层提取到的特征图输入到ABM中,有效提取更丰富的边缘特征信息.最后,将主分支和辅助分支的结果融合作为输出,融合后的输出不仅追踪到了不同尺度的空间特征信息,而且提取到了完整的边缘特征信息,从而使模型更有效地提高分割精度.PASCAL VOC 2012数据集的结果表明,改进后的模型相比于原模型分割精度提升了3.21%,与其它网络模型相比,也具有较好的分割精度.An improved DeepLabv3 model for semantic segmentation of natural images is proposed in this paper.The problem of local detail information loss that leads to mis-segmentation and incomplete object edge segmentation in the DeepLabv3 model is addressed.Higher accuracy in semantic segmentation of natural images can be achieved.Firstly,ResNet101 is used as the backbone network for feature extraction.The feature maps extracted from the last two layers of ResNet101 are input into the designed ACMix multiple fusion module(AMFM)to effectively obtain spatial feature information at different scales.The input for the Atrous spatial pyramid pooling(ASPP)module is comprised of the fused results.Secondly,an Auxiliary branch module(ABM)is introduced,where the feature maps extracted from the third layer of ResNet101 are input.More abundant edge feature information is extracted effectively.Finally,the output is formed by merging the results of the main branch and the auxiliary branch.The fused output not only tracks spatial feature information at different scales but also extracts complete edge feature information,thereby significantly enhancing the segmentation accuracy of the model.The results from the PASCAL VOC 2012 dataset demonstrate that the segmentation accuracy of the improved model is improved by 3.21%compared to the original model.Furthermore,the improved model exhibits good segmentation accuracy compared to other network models.

关 键 词：语义分割 DeepLabv3 多尺度特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]