顾及小尺度目标特征重建的全局语义分割模型  

Global Semantic Segmentation Model Considering Small-scale Target Feature Reconstruction

作　　者：吴小所[1] 乔煜栋 贺成龙 刘小明 闫浩文 WU Xiaosuo;QIAO Yudong;HE Chenglong;LIU Xiaoming;YAN Haowen(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Engineering School,Qinghai Institute of Technology,Xining 810016,China)

机构地区：[1]兰州交通大学电子与信息工程学院,甘肃兰州730070 [2]青海理工学院工学院,青海西宁810016

出　　处：《湖南大学学报(自然科学版)》2025年第4期44-56,共13页Journal of Hunan University:Natural Sciences

基　　金：国家重点研发计划资助项目(2022YFB3903604);甘肃省自然科学基金资助项目(21JR7RA310);兰州交通大学青年科学基金资助项目(2021029)。

摘　　要：针对复杂背景下航空遥感图像中多类别小尺度目标特征的理解困难和特征边界分割不清晰的问题,本研究构建了一种新型的分割模型,该模型通过综合主干网络特征并进行特征分类与重构来提升分割效果.模型以Swin-Transformer作为基础编码结构,利用其强大的全局语义信息捕捉能力进行特征抽取.进一步,本研究创新性地提出了信息聚合重构模块(IGRM)和通道区分重构模块(CRRM),这两种结构能够依据信息量对抽取的特征进行分类和重构,以此细化了对小尺度目标特征的处理.模型结合了上采样与下采样的特征连接,并将重构特征与编码器特征融合,形成多尺度特征聚合块,进而输出精确的分割结果.在处理复杂背景下的多目标场景时,本模型能够对细小尺度目标特征进行精确重构,生成高分辨率的分割图像,显著提升了分割的准确度.在ISPRS Potsdam和ISPRS Vaihingen数据集上,本模型取得了平均交并比(mIoU)分别为87.15%和82.93%、整体准确率(OA)分别为91.53%和91.4%的优异表现.为评估模型对多类别小尺度目标特征提取的泛化性能,本文还进行了针对复杂背景下小车类别的对比实验,在UAVid数据集上的mIoU达到了67.86%.To solve the problems of insufficient understanding and fuzzy feature boundary segmentation of multi-target categories small-scale feature semantic information in aerial remote sensing images in complex background,this paper designs a segmentation model that integrates the features of the backbone network information and classifies and reconstructs the features to improve the segmentation effect.The model takes Swin-Transformer as the coding structure and utilizes its ability to understand global semantic information for feature extraction.The segmentation of small-scale target features is refined by the designed information grouping reconstruction convolution(IGRM)and channel classification reconstruction convolution(CRRM),which classify and reconstruct the extracted features by the amount of information.Finally,by integrating the up-sampling and down-sampling connections,the reconstructed features are fused with the features extracted by the encoder to form a multi-scale feature aggregation block to output the segmentation results.The refined reconstruction of small-scale target features is realized in multi-target scenarios with complex backgrounds,and high-quality segmentation maps are generated to improve the segmentation accuracy.Experimental results on the ISPRS Potsdam and ISPRS Vaihingen datasets show that the average intersection and merger ratio(mIoU)is 87.15%and 82.93%,respectively,and the overall accuracy(OA)is 91.53%and 91.4%,respectively.To verify the generalization ability of the model for small-scale target feature extraction in multi-target categories,this paper also designs a comparative experiment for the category of carts in complex backgrounds.The experimental results show that the mIoU on the UAVid dataset reaches 67.86%.

关 键 词：航空遥感 语义分割 信息聚合重构模块 通道区分重构模块 整合上采样 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] TP391.41[自动化与计算机技术—控制科学与工程]