摘要
卷积神经网络(Convolutional Neural Networks,CNN)在处理语义分割上具有独特优势,但是小目标、稀少目标等样本不均衡问题成为影响分割精度的重要因素。由此提出一种以DeepLabv3网络为基本框架,结合了Focal Loss和GPB/OWT/UCM边缘分割的图像语义分割方法。首先,利用DeepLabv3网络中Resnet101编码器进行特征提取,所得特征经过空洞空间金字塔池化(Atrous Spatial Pyramid Pooling,ASPP)模块后获得多尺度信息,再通过解码器恢复到原图分辨率,进行图像的像素级分割;然后,引入Focal Loss替换原来网络中的交叉熵损失函数,即在原来每个单像素损失值上加权重,加强对困难样本特征的学习;最后,融合GPB/OWT/UCM边缘分割算法,使得小目标区域的语义信息仅由低等级特征去单独决定,从而优化小目标和稀少目标的边缘细节的分割。实验结果表明:在PASCAL VOC 2012数据集上,该方法所得均交并比(Mean Intersection over Union,MIoU)为80.23%,相比于原来DeepLabv3网络,MIoU值提升了0.71%,改善了样本不均衡问题。
Semantic segmentation is a pixel-level classification task.Convolutional Neural Networks(CNN)have unique advantages in processing semantic segmentation,but the problem of imbalanced samples,such as small targets and rare targets have become an important facor affecting the accuracy of segmentation.In response to this problem,an semantic image segmentation method based on DeepLabv3 network was proposed,which combined with Focal Loss and GPB/OWT/UCM edge segmentation.First,the Resnet101 encoder structure in the DeepLabv3 network was used for feature extraction,and the obtained features were subjected to the Atrous Spatial Pyramid Pooling(ASPP)module to get multi-scale information,and then restorted to the original image resolution through the decoder,and the image pixel-level segmentation was performed;then,the Focal Loss was introduced to replace the cross-entropy loss function in the original network,that is,adding weights to each original single-pixel loss value to strengthen the learning of difficult sample features;Finally,the GPB/OWT/UCM edge segmentation algorithm was integrated,so that the semantic information of the small target area was determined solely by low-level features,thereby optimizing the edge details segmentation of small targets and rare targets.Experimental results show that,on the PASCAL VOC 2012 data set,Mean Intersection over Union(MIoU)obtained by this method is 80.23%.Compared with the original DeepLabv3 network,the MIoU value has increased by 0.71%.Obviously,this method improve the sample imbalance problem.
作者
王汉谱
刘志豪
谷旭轩
廖建英
贺志强
涂兵
彭怡书
WANG Hanpu;LIU Zhihao;GU Xuxuan;LIAO Jianying;HE Zhiqiang;TU Bing;PENG Yishu(School of Information and Communication Engineering,Hunan Institute of Science and Technology,Yueyang 414006,China;Machine Vision&Artificial Intelligence Research Center,Hunan Institute of Science and Technology,Yueyang 414006,China)
出处
《成都工业学院学报》
2022年第3期16-21,共6页
Journal of Chengdu Technological University
基金
国家自然科学基金项目(61977022)
湖南省自然科学基金项目(2020JJ4340)
湖南省研究生教育创新工程和专业能力提升工程项目(CX20201114)
湖南省教育厅优秀青年项目(19B245)
湖南省三维重建与智能应用技术工程研究中心(2019-430602-73-03-006049)
湖南省应急通信工程技术研究中心(2018TP2022)
2021年湖南省大学生创新训练项目(基于图像语义分割的洞庭湖沙船目标识别研究)。
关键词
语义分割
样本不均衡
多尺度信息
边缘细节
semantic segmentation
sample imbalance
multi-scale information
edge detail
