基于ASPP模块及特征图加权融合的图像语义分割算法  

IMAGE SEMANTIC SEGMENTATION ALGORITHM BASED ON ASPP MODULE AND WEIGHTED FUSION OF FEATURE MAP

作　　者：祁新 杜洪波[1] 朱立军[2] Qi Xin;Du Hongbo;Zhu Lijun(School of Science,Shenyang University of Technology,Shenyang 110870,Liaoning,China;School of Information and Computing Science,Northern University for Nationalities,Yinchuan 750021,Ningxia,China)

机构地区：[1]沈阳工业大学理学院,辽宁沈阳110870 [2]北方民族大学信息与计算科学学院,宁夏银川750021

出　　处：《计算机应用与软件》2025年第3期227-232,267,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(11861003);辽宁省教育厅高等学校基本科研项目(LJKZ0157)。

摘　　要：针对现有语义分割算法对特征图进行高级语义信息提取时存在部分低、中层级细节信息丢失的问题,提出基于ASPP模块及特征图加权融合的图像语义分割算法。改进算法借助“残差思想”,建立特征提取网络与ASPP模块之间的跳跃连接,融合特征图加权,增加ASPP模块提取图像语义信息的能力,减少低、中层部分细节信息丢失,并在相应卷积层后进行批归一化,缓解梯度消失的问题,提高模型的分割性能和运行效率。实验结果表明,特征图的合理融合使得语义分割图边缘细节更好地保留,改进算法分割精度提高,其平均交并比比原模型提高6.8%,频率加权交并比提高5.33%。Aimed at the problem that some low-level and middle-level detail information is lost when the existing semantic segmentation algorithms extract high-level semantic information from feature map,an image semantic segmentation algorithm based on ASPP module and weighted fusion of feature map is proposed.With the help of the“residual idea”,the improved algorithm established the jump connection between the feature extraction network and the ASPP module,integrated the weighting of the feature map,increased the ability of the ASPP module to extract image semantic information,reduced the loss of detail information in the low-level and middle-level layers,and performed batch normalization after the corresponding convolution layer to alleviate the problem of gradient disappearance and improve the segmentation performance and operation efficiency of the model.The experimental results show that the reasonable fusion of feature maps makes the edge details of semantic segmentation map better preserved,and the segmentation accuracy of the improved algorithm is enhanced.The average intersection and union ratio is 6.8%higher than the original model,and the frequency weighted intersection and union ratio is 5.33%.

关 键 词：语义分割 ASPP模块 特征图加权 平均交并比 频率加权交并比 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]