多尺度融合卷积的轻量化Transformer无人机地物识别模型  

A Lightweight Transformer UAV Surface Feature Recognition Model Based on Multi-scale Fusion Convolutional Networks

作　　者：肖斌[1] 罗浩 张恒宾 刘宏伟 张兴鹏 XIAO Bin;LUO Hao;ZHANG Hengbin;LIU Hongwei;ZHANG Xingpeng(School of Computer Science,Southwest Petroleum University,Chengdu 610500,China)

机构地区：[1]西南石油大学计算机科学学院,四川成都610500

出　　处：《郑州大学学报（理学版）》2024年第1期32-39,共8页Journal of Zhengzhou University:Natural Science Edition

基　　金：国家自然科学基金项目(62006200);四川省自然科学基金项目(2022YFG0179);油气藏地质与开发国家重点实验室开放基金项目(成都理工大学)(PLC20211104)。

摘　　要：Transformer模型性能优越,但其巨大的参数量不适合资源受限的无人机遥感任务。为此,提出一种用于无人机遥感图像的多尺度融合卷积的轻量化Transformer模型,通过设计三种优化策略来提高精度以及减少参数量。首先,设计了一种轻量级多尺度融合卷积方法,补充Transformer丢失的块内空间信息,从而有效提取多尺度上的粗、细粒度特征表示。其次,设计了多尺度缩减键值序列的方式,优化Transformer中的自注意力计算。最后,设计了轻量级的MLP解码器,进一步减少模型参数量。在Vaihingen和Potsdam数据集上与一些主流模型进行了对比实验,结果表明,所提模型的F 1值和交并比均有所提升。同时,在Potsdam数据集上准确度提升0.29%,参数量比双分支网络STransFuse减少18%。The Transformer model had excellent performance,but its tremendous number of parameters made it unsuitable for UAV remote sensing mission with limited resources.Therefore,a lightweight Transformer model for multi-scale fusion convolution networks of UAV remote sensing images was proposed,and three optimization strategies were designed to improve the accuracy and reduce the number of model parameters.Firstly,a lightweight multi-scale fusion convolution method was designed to supplement the intra-block spatial information lost by Transformer,so as to effectively extract the coarse and fine grained feature representation at multiple scales.Secondly,a multi-scale key-value sequence reduction method was devised to optimize the self-attention calculation in Transformer.Finally,a lightweight MLP decoder was applied to further reduce the number of model parameters.The comparative experiments with some mainstream models on Vaihingen and Potsdam data sets showed that the F 1 value and intersection over union of the proposed model were improved.Meanwhile,the accuracy of the Potsdam data set was improved by 0.29%,and the parameters were reduced by 18%compared with the dual branch network STransFuse.

关 键 词：无人机遥感影像 TRANSFORMER 语义分割 轻量级 多尺度 卷积神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]