机载广域遥感图像的尺度归一化目标检测方法  

Multi-scale normalized detection method for airborne wide-area remote sensing images

作　　者：朱圣杰 王宣[1] 徐芳[1] 彭佳琦 王远超 ZHU Sheng-jie;WANG Xuan;XU Fang;PENG Jia-qi;WANG Yuan-chao(Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China;Daheng College,University of Chinese Academy of Sciences,Beijing 100049,China;First Military Representative Office in Changchun,Changchun 130033,China;Shanghai Electro-Mechanical Engineering Institute,Shanghai 201109,China)

机构地区：[1]中国科学院长春光学精密机械与物理研究所,长春130033 [2]中国科学院大学大珩学院,北京100049 [3]驻长春地区第一军事代表室,长春130033 [4]上海机电工程研究所,上海201109

出　　处：《吉林大学学报（工学版）》2024年第8期2329-2337,共9页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(No.61905240).

摘　　要：针对机载广域遥感图像的目标尺寸变化大、背景噪声复杂以及局部目标密集给目标检测任务带来的困难,本文通过优化分割方法统一输入图像的目标像素尺寸,并以此简化模型结构提出了一种尺度归一化卷积神经网络模型MNNet。为增强局部之间的特征关联,本文设计了全局连接块(SGC),有效提高了检测的精度。针对现有非极大值抑制算法的超参数依赖经验设置的问题,本文提出了一种自适应非极大值抑制方法(DNMS),降低了模型的部署难度。在RSF数据集上的测试结果表明:本文模型的检测平均精度(AP)高于其他模型5.0%以上,在检测速度上达到了57.7帧/s,可以满足遥感图像的检测任务需求。Aiming at the difficulty of object detection caused by the large target size variation,complex background noise and dense targets in airborne wide-area remote sensing images,this paper unifies the target pixel size of the input image by optimizing the segmentation method,and proposes a multi-scale normalized convolutional neural networks model(MNNet).To enhance the feature correlation between localities,this paper designs a space global connection block(SGC),which effectively improves the detection accuracy.For the problem that the parameters of the existing NMS algorithm depend on the empirical setting,this paper proposes a self-adaption non-maxima suppression method(DNMS),which reduces the difficulty of model deployment.The test results on the RSF dataset show that the average precision(AP)of the model in this paper is higher than that of other models by more than 5.0%,and the detection speed reaches 57.7 fps,which can meet the detection task of remote sensing images.

关 键 词：模式识别与智能系统 计算机视觉 目标检测 遥感图像 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]