基于Deeplabv3+的街景图像语义分割技术研究  

Research on semantic segmentation of streetscape image based on deep labv3+

作　　者：董淼[1] 牛蔓丽 蒋纬昌 Dong Miao;Niu Manli;Jiang Weichang(Henan Polytechnic Institute,architectural engineering institute,HeNan,NanYang 473000;China Mobile Communications Group Henan Co.,Ltd.Nanyang Branch,HeNan,NanYang 473000)

机构地区：[1]河南工业职业技术学院建筑工程学院,河南南阳473000 [2]中国移动通信集团河南有限公司南阳分公司,河南南阳473000

出　　处：《现代科学仪器》2023年第2期160-165,共6页Modern Scientific Instruments

基　　金：2022年度河南省重点研发与推广专项科技攻关项目“基于大数据开发和机器学习算法的智慧旅游技术研发”,项目编号:222102320115。

摘　　要：随着城市化的发展,城市街道场景的构成越来越复杂,Deeplabv3+街景语义分割模型及其常见的几种改进模型已经难以满足现实需求。为了提高模型的性能和速度,改进了模型的特征提取网络,用改进GAN网络代替Xception网络,并在输入原始图像时同时输入灰度图像,以Sobel算子边缘检测提取灰度特征;此次研究还改进了模型的解码器,添加CBAM-SE模块,生成图像的多维特征信息,使用交叉熵损失函数降低内部加权;最后对模型的各模块组合方案进行消融实验,详细分析了各模块对模型整体性能的提升作用。经过训练和测试,此模型的分割精度相较于常见的几种改进模型提升了0.39%-0.7%。With the development of urbanization,the composition of urban street scenes is becoming more and more complex.Deeplabv3+street scene semantic segmentation model and several common improved models have been difficult to meet the practical needs.In order to improve the performance and speed of the model,a GaN algorithm based on RDB structure and Sobel operator for edge detection is designed in this study.RDB structure is used to replace RB structure,and gray image input is added in the input module.Sobel operator is used as the edge detection operator;The cbam-se module is added to the decoder to generate the multi-dimensional feature information of the image.Finally,the cross entropy loss function is used to reduce the internal weighting to improve the performance and speed of the model.After experiments and tests,compared with several common improved models,the segmentation accuracy of this model is improved by 0.39%-0.7%,which can meet the requirements of complex street scene semantic segmentation.

关 键 词：Deeplabv3+ GAN RDB CBAM-SE 复杂街景 语义分割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]