融合注意力机制的改进型DeepLabv3+语义分割  

Improved DeepLabv3+semantic segmentation incorporating attention mechanisms

作　　者：闫河 雷秋霞 王旭 YAN He*;LEI Qiuxia;WANG Xu(Liangjiang College of Artificial Intelligence,Chongqing University of Technology,Chongqing 401135,China)

机构地区：[1]重庆理工大学两江人工智能学院,重庆401135

出　　处：《光学精密工程》2025年第1期123-134,共12页Optics and Precision Engineering

基　　金：国家重点研发计划“智能机器人”重点专项项目(No.2018YFB1308602);国家自然科学基金面上项目(No.61173184);重庆市自然科学基金项目(No.cstc2018jcy-jAX0694)。

摘　　要：针对DeepLabv3+语义分割网络计算复杂度高、对图像细节提取能力弱、分割的图像边界模糊的问题,提出了一种融合注意力机制的改进型DeepLabv3+语义分割网络。以轻量级网络MobileNetV2为骨干,在保持较高表征能力的同时显著减少模型参数,在骨干网络的低层特征后面加入轻量级、无参数注意力机制(Simple,Parameter-Free Attention Module,SimAM),对输入的特征进行加权,以增强关键特征的提取能力。将ASPP模块的全局平均池化替换成Haar小波变换下采样(Haar Wavelet Downsampling,HWD),以避免丢失空间信息,同时在ASPP模块之后加入外部注意力机制(External Attention,EANet),以更好地利用上下文信息,实现多尺度融合,从而提升语义理解能力和语义分割的准确性。实验结果表明,该模型在VOC2012数据集上相较于原有的DeepLabv3+语义分割模型,平均交并比(mIoU)提高了2.82%。本文提出的改进模型显著提高了模型语义分割的精度,为计算机视觉领域应用提供了新的思路。To address the challenges of high computational complexity,limited detail extraction,and fuzzy boundaries in the current DeepLabv3+semantic segmentation network,this study proposes an en-hanced DeepLabv3+model incorporating attention mechanisms.Specifically,the lightweight Mobile-NetV2 is employed as the backbone to balance high representational capacity with a significant reduction in model parameters.A parameter-free lightweight attention mechanism(SimAM)is integrated into the low-level features of the backbone network to prioritize key features and enhance feature extraction capabilities.Furthermore,the global average pooling in the ASPP module is replaced with Haar Wavelet Transform Downsampling(HWD)to preserve spatial information.An External Attention Mechanism(EANet)is al-so introduced after the ASPP module to leverage contextual information and achieve multi-scale feature fu-sion,thereby improving semantic understanding and segmentation accuracy.Experimental results demon-strate that the proposed model achieves a 2.82%improvement in mean Intersection over Union(mIoU)on the VOC2012 dataset compared to the original DeepLabv3+model.This research enhances the preci-sion of semantic segmentation and offers novel insights for advancing applications in computer vision.

关 键 词：语义分割 DeepLabv3+ Haar小波变换下采样 外部注意力机制 多尺度融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP181[自动化与计算机技术—计算机科学与技术]