改进ASPP和融合多尺度特征语义分割算法  

Improved ASPP and Semantic Segmentation Algorithm with Fused Multi-scale Features

作　　者：曹新宇 张太红[1,2,3] 赵昀杰 CAO Xin-yu;ZHANG Tai-hong;ZHAO Yun-jie(College of Computer and Information Engineering,Xinjiang Agricultural University,Urumchi Xinjiang 830000,China;Engineering Research Center of Intelligent Agriculture,Ministry of Education,Urumchi Xinjiang 830000,China;Xinjiang Agricultural Informatization Engineering Technology Research Center,Urumchi Xinjiang 830000,China)

机构地区：[1]新疆农业大学计算机与信息工程学院,新疆乌鲁木齐830000 [2]智能农业教育部工程研究中心,新疆乌鲁木齐830000 [3]新疆农业信息化工程技术研究中心,新疆乌鲁木齐830000

出　　处：《计算机仿真》2025年第3期208-215,476,共9页Computer Simulation

基　　金：科技创新2030—“新一代人工智能”重大项目(2022ZD0115805);新疆维吾尔自治区重大科技专项(2022A02011)。

摘　　要：针对语义分割算法DeepLabv3+,并未对骨干网络提取的特征信息充分利用,导致最后的分割边缘不连续、模糊以及分割错误等问题,提出了改进ASPP和融合多尺度特征的语义分割算法。算法采用极化自注意力和全注意力空洞空间金字塔池化模块(FP-ASPP)建立起多个信息流之间的交互,增强信息之间的联系,捕获较为密集的像素点,更加充分的利用深层的语义信息。其次引入了特征对齐金字塔模块(Feature Aligned Pyramid Module),对骨干网络的多层特征信息进行利用,增强特征的表达能力,利用其中的FAM和FSM对不同尺度特征信息进行融合,提高各层特征之间的互补能力,来获取到更加全面的特征图。最后将FP-ASPP和特征对齐模块的输出进行拼接操作,得到最后的分割结果。在IDD数据集和Cityscapes数据集上进行大量实验,结果表明所提出的算法的平均交并比(mIoU)分别达到65.78%和66.64%,相较于基准算法DeepLabv3+分别提升了1.33%和2.17%。该算法充分利用了骨干网络中的每层特征信息,在原有的算法基础上提升了分割精度,也取得了更好的分割结果。Aiming at the semantic segmentation algorithm DeepLabv3+,which does not fully utilize the feature information extracted by the backbone network,resulting in discontinuous,blurred,and erroneous segmentation edges,an improved ASPP and multi-scale features fusion semantic segmentation algorithm is proposed.The algorithm uses polarized self-attention and full-attention null space pyramidal pooling module(FP-ASPP)to establish the interaction between multiple information streams,enhance the connection between information,capture denser pixel points,and more fully utilize the deep semantic information.Secondly,the Feature Aligned Pyramid Module(FAM)is introduced to utilize the multi-layer feature information of the backbone network to enhance the feature representation capability,and use the FAM and FSM to fuse the feature information at different scales to improve the complementary capability between the features of each layer to obtain a more comprehensive feature map.Finally,the outputs of FP-ASPP and feature alignment modules are stitched together to obtain the final segmentation results.The results of extensive experiments on IDD dataset and Cityscapes dataset show that the average intersection-to-merge ratio(mloU)of the proposed algorithm reaches 65.78% and 66.64%,respectively,which is 1.33% and 2.17% better than the benchmark algorithm DeepLabv3+,respectively.The algorithm makes full use of the feature information of each layer in the backbone network to improve the segmentation accuracy based on the original algorithm,and also achieves better segmentation results.

关 键 词：语义分割 空洞空间金字塔池化 特征对齐金字塔 特征融合 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]