福州大学学报（自然科学版）2026，Vol.54Issue(2)：137-144,8.DOI:10.7631/issn.1000-2243.25088
融合多层级特征与注意力机制的高效语义分割算法
An efficient semantic segmentation algorithm fusing multi-level features and attention mechanism
摘要
Abstract
Due to the current semantic segmentation algorithm's numerious parameters,complex calculations,slow inference speed,it is difficult to apply in resource-limited scenarios such as mobile devices and embedded devices.Soan efficient semantic segmentation algorithm named HE-DeepLabV3+that integrates multi-level features and attention mechanism is proposed by improving DeepLabV3+.First,MobileNetV2 is used as the backbone network,and the output shallow features are fused with multi-level features to enhance the expressive ability of shallow features by the feature pyramid networks(FPN)module.Afterwards,the enhanced atrous spatial pyramid pooling(EASPP)module is proposed to realize dense connection of features and to prevent semantic information loss in the feature extraction process.Then the parallel channel-spatial attention module(PCSAM)module is designed to obtain spatial weights and channel weights while retaining the original features,and to improve the segmentation accuracy.Experimental results show that the average intersection ratio and average pixel accuracy of this algorithm on the PASCAL VOC2012Aug data set reached 75.44%and 84.99%.By comparing this algorithm with the traditional DeepLabV3+,the amount of parameters and calculations reduced by 82.55%and 62.52%respectively,meanwhile the inference speed increased by 71.92%.This algorithm significantly reduces computational costs while maintaining high accuracy,achieving an effective balance between accuracy and model lightweight.关键词
语义分割/DeepLabV3+/注意力机制/多层级特征融合Key words
semantic segmentation/DeepLabV3+/attention mechanism/multi-level feature fusion分类
信息技术与安全科学引用本文复制引用
郑仕敏,毕建鹏,于潇雁..融合多层级特征与注意力机制的高效语义分割算法[J].福州大学学报（自然科学版）,2026,54(2):137-144,8.基金项目
科技部国家重点研发计划资助项目(2022YFB4702401) （2022YFB4702401）
福建省自然科学基金资助项目(2024J01241) （2024J01241）
