基于显著性语义协同的校园道路场景分割  

Campus scene segmentation based on salient semantic collaborative

作　　者：刘蓝蓝[1,2] 杜敏敏 郑伟 司马海峰 LIU Lanan;DU Minmin;ZHENG Wei;SIMA Haifeng(School of Software,Henan Polytechnic University,Jiaozuo 454000,Henan,China;Faculty of Arts and Law,Henan Polytechnic University,Ji-aozuo 454000,Henan,China)

机构地区：[1]河南理工大学软件学院,河南焦作454000 [2]河南理工大学文法学院,河南焦作454000

出　　处：《河南理工大学学报（自然科学版）》2023年第4期148-155,共8页Journal of Henan Polytechnic University（Natural Science）

基　　金：国家自然科学基金资助项目(61602157);河南省科技攻关计划项目(202102210167);河南省高校创新团队项目(19IRTSTHN012);河南理工大学博士基金资助项目(B2016-37)。

摘　　要：针对复杂场景图像语义分割精度低、细小目标识别能力不足等问题,提出基于显著性语义协同学习的校园道路场景分割算法,并在实时网络模型LinkNet基础上对该算法加以改进。首先,在编码阶段引入空洞空间金字塔池化模块,通过扩大视野提取更丰富的上下文语义信息,提升细小目标的分割能力;其次,采用协同学习思想,将语义分割与显著性检测的特征提取过程进行共享。通过共享卷积层特征,语义分割任务从显著性的学习过程中受益,进而提高分割模型的准确率。为了验证算法性能,在Cityscapes数据集上进行实验。结果表明,与经典的语义分割方法进行对比,本文算法能进一步提高场景内各类目标分割精确度,整体精度达到67.91%,比原LinkNet模型提高了8.14%。Aimed at low segmentation accuracy and insufficient recognition of small targets,a semantic seg-mentation algorithm based on salience collaborative was proposed,it was an improved model of the real-time network LinkNet.Firstly,the Atrous Spatial Pyramid Pooling was introduced to enlarge receptive field.This measure was able to enrich the context semantic information,thus the segmentation ability of small tar-gets could be improved.Secondly,collaborative learning was adopted to share feature extraction process.The processes included semantic segmentation and saliency detection.The segmentation task benefited from train-ing of saliency by sharing the features of convolution layer,then the accuracy of this model was improved.The proposed model was trained on the Cityscapes dataset to verify the performance.Experiment results showed that the proposed algorithm improved the accuracy of various targets of scene to 67.91%,which was 8.14%higher than the original LinkNet model.

关 键 词：校园安全 图像语义分割 协同学习 显著性语义 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]