基于大核卷积和密集目标细化的遥感图像多尺度特征增强网络  

Multi⁃Scale Feature Enhancement Network Based on Large Kernel Convolution and Dense Object Refinement for Remote Sensing Images

作　　者：王占魁 秦品乐[1] 曾建潮 WANG Zhankui;QIN Pinle;ZENG Jianchao(School of Computer Science and Technology,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学计算机科学与技术学院,山西太原030051

出　　处：《中北大学学报（自然科学版）》2024年第5期628-637,共10页Journal of North University of China(Natural Science Edition)

基　　金：山西省科技重大专项计划“揭榜挂帅”项目(202101010101018)。

摘　　要：针对遥感图像中目标尺度变化差异大、方向任意和分布密集,现有检测方法较少直接关注密集边缘信息且目标无法获得合适的感受野,遥感检测效果较差的问题,本文提出了一种基于大核卷积和密集目标细化的多尺度特征增强网络(LKCSFP-NET)来进行遥感图像的检测。该网络首先在SKNET基础上增加了空洞卷积形成大核卷积块(LKB),从而获得小目标的最佳感受野以及提升了网络对多尺度的适应性和准确度;其次在FPN基础上增加了集中空间特征金字塔CSFP模块,通过将全局语义信息与局部语义信息相结合,解决了遥感图像因目标分布密集以及背景复杂导致的检测效率较低的问题。实验结果表明,在DOTA和HRSC2016公开数据集上,所提算法在2个数据集上的平均检测精度分别为75.96%和96.60%,较基线网络提升了1.36百分点和0.63百分点,优于现有大多数模型。所提出的LKCSFP-NET在两个公开数据集中表现稳定,对小目标和密集排列的目标都有较好的检测效果,高于现有大多数模型的检测精度,可以很好地应用于遥感目标的检测。Due to the large difference in the scale of the object,the arbitrary direction and the dense distribution of the object in the remote sensing image,the existing detection methods rarely pay direct attention to the dense edge information and the object cannot obtain a suitable receptive field,so it is difficult to have good detection results in remote sensing detection.In order to solve the above problems,this paper proposed a multi-scale feature enhancement network based on large kernel convolution and dense object refinement(LKCSFP-NET)for remote sensing image detection.Firstly,the network based on SKNET added a cavity convolution to form a large kernel convolution block(LKB)to obtain the best sensitivity field for small targets and improve the adaptability and accuracy of the network to multiple scales.Secondly,on the basis of FPN,the centralized spatial feature pyramid CSFP module was added to solve the problem of low detection efficiency of remote sensing images due to dense object distribution and complex detection background by combining global semantic information with local semantic information.The experimental results show that on the DOTA and HRSC2016 public datasets,the average detection accuracy of the proposed algorithm on the two datasets is 74.90%and 96.60%,respectively,which is 1.36 and 0.63 percentage points higher than that of the baseline network,which is better than most existing models.The proposed LKCSFP-NET has stable performance in the two public datasets,and has good detection results for small objects and densely arranged objects,which is higher than the detection accu‐racy of most existing models,and can be well applied to the detection of remote sensing objects.

关 键 词：目标检测 遥感图像 多尺度 大核卷积 密集检测 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP242[自动化与计算机技术—计算机科学与技术]