基于特征重用和膨胀卷积的遥感图像舰船检测  

Ship detection in remote sensing images via feature reuse and dilated convolution

作　　者：曲海成[1] 李瑞柯 王蒙 单以盟 QU Haicheng;LI Ruike;WANG Meng;SHAN Yimeng(College of Software,Liaoning Technical University,Huludao 125105,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125105

出　　处：《智能系统学报》2024年第5期1298-1308,共11页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金面上项目(42271409);辽宁省高等学校基本科研项目(LIKMZ20220699).

摘　　要：在光学遥感图像中,港口内的舰船目标通常处于密集的船只群中,并受到周围环境的干扰和遮挡,如集装箱、车辆等。为了进一步提高现有舰船目标检测算法的精度和泛化性能,提出了一种基于特征重用和膨胀卷积的遥感图像舰船检测算法。首先构建了基于分组卷积和拆分注意力的残差块来提取特征,同时嵌入可变形卷积提取更加符合舰船尺度变化的特征;接着,构造了多尺度感受野模块,通过并行提取多尺度特征后再进行融合来减少信息损失;最后,在原有特征金字塔的基础上构建了一条自底向上的特征重用聚合路径以提高特征表示能力。在大型遥感数据集DOTA和舰船数据集HRSC2016上进行实验,实验结果表明,所提方法能够有效缓解舰船目标漏检和误检问题,提高了遥感图像舰船目标检测的精度。In optical remote sensing images,ship targets in ports are often densely grouped and obstructed by the surrounding environment,such as containers and vehicles.To further improve the accuracy and generalization performance of existing ship target detection algorithms,this study proposes a remote sensing image ship detection algorithm based on feature reuse and dilated convolution.First,a residual block based on grouped convolution and split attention is constructed to extract features,with deformable convolution embedded to better handle ship scale variations.Afterward,a multiscale receptive field module is designed to reduce information loss by parallel extraction and fusion of multiscale features.Finally,a bottom-up feature reuse aggregation path is developed based on the original feature pyramid to enhance feature representation.Experiments were conducted on the large-scale remote sensing dataset,DOTA and the ship dataset,HRSC2016.The results show that the proposed method effectively alleviates the issues of missing and false detections of ship targets,leading to increased accuracy in ship target detection in remote sensing images.

关 键 词：遥感图像 舰船检测 特征重用 膨胀卷积 拆分注意力 分组卷积 特征金字塔 可变形卷积 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置] TN911.73[自动化与计算机技术—控制科学与工程]