改进STDC-Seg的实时图像语义分割网络算法  

Real-time image semantic segmentation network algorithm based on improved STDC-Seg

作　　者：兰建平 董冯雷 杨亚会 董秀娟 胡逸贤 LAN Jianping;DONG Fenglei;YANG Yahui;DONG Xiujuan;HU Yixian(Institute of Automotive Engineers,Hubei University of Automotive Technology,Shiyan 442002,China)

机构地区：[1]湖北汽车工业学院汽车工程师学院,湖北十堰442002

出　　处：《传感器与微系统》2023年第11期110-113,118,共5页Transducer and Microsystem Technologies

基　　金：2023年度湖北汽车工业学院教学研究与改革项目创新创业教育专项项目(SCJY202314);2022年湖北省教育厅科学研究计划指导性项目(B2022397)。

摘　　要：针对目前自动驾驶等领域下语义分割算法实时性不足的问题,提出一种基于编码器—解码器结构的轻量级实时语义分割网络模型。首先,使用短期密集级联(STDC)网络作为骨干网络用以提取图像特征;然后,构建一种轻量级解码器,以减少解码器部分的计算开销;最后,为加强解码器阶段的特征表达,提出2种注意力融合模块对不同层次的特征进行加权融合。实验结果表明:该算法在Cityscapes测试集上实现了74.22%mIoU和154.7 fps的性能;与STDC-Seg网络相比,该算法减少了21.3%的参数量,平均交并比(mIoU)高出2.02个百分点。在实时性和准确性方面获得了很好的平衡。Aiming at the problem of insufficient real-time of semantic segmentation algorithms in domains such as autonomous driving,a lightweight real-time semantic segmentation network model based on encoder-decoder structure is proposed.Firstly,a short-term dense cascade(STDC)network is used as backbone network to extract image features.Then,a lightweight decoder is built to reduce the computational overhead of the decoder part.Finally,two attention fusion modules are proposed to enhance the feature representation in the decoder stage to weight and fuse the features at different levels.Experimental results show that the algorithm achieves a performance of 74.22% mIoU and 154.7 fps on the Cityscapes test set.Compared with the STDC-Seg network,the algorithm reduces the number of parameters by 21.3%,and the mean intersection over union(mIoU)is 2.02 percentage points higher.A good balance is obtained in terms of real-time performance and accuracy.

关 键 词：实时语义分割 轻量化卷积神经网络 深度学习 注意力机制 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]