基于分数阶Gabor变换卷积的遥感图像飞机目标检测算法  

Aircraft Target Detection Algorithm in Remote Sensing Image Based on Fractional Gabor Transform Convolution

作　　者：陈昕 单慧琳 段修贤 吴心悦 马丁 张银胜 CHEN Xin;SHAN Huilin;DUAN Xiuxian;WU Xinyue;MA Ding;ZHANG Yinsheng(School of Electronics and Information Engineering,Wuxi University,Wuxi 214105,China)

机构地区：[1]无锡学院电子信息工程学院,无锡214105

出　　处：《航天返回与遥感》2025年第1期135-149,共15页Spacecraft Recovery & Remote Sensing

基　　金：国家自然科学基金项目(62071240,62106111);国家级大创项目(202413982004Z)。

摘　　要：针对遥感图像飞机目标检测任务所面临的背景噪声、目标尺寸和旋转角度等因素干扰特征提取的问题,文章提出了一种基于分数阶Gabor变换卷积的遥感图像飞机目标检测算法。首先,在特征提取网络将高效层聚合网络和卷积块注意力模块结合,形成一种新型高效的注意力特征提取模块。其次,在特征融合网络构建分数阶Gabor变换卷积模块,通过突出飞机目标的边缘、纹理及方向等细节特征信息来改善特征融合效果。最后,在检测层采用可学习动态检测头,通过尺度感知注意力模块加强对多尺度目标的关注、通过空间感知注意力模块加强辨别空间位置、通过任务感知注意力模块更准确地区分任务需求。在DOTAv1数据集上进行的实验结果表明,文章方法检测精度达96.2%,相较基线模型YOLOv7提升了2.2个百分点,模型权重更小,在复杂场景下的检测精度提升明显。该方法为遥感图像飞机目标检测提供了一种更高效的检测方案。This paper proposes a fractional-order Gabor transform convolution-based algorithm for aircraft target detection in remote sensing images,addressing challenges such as background noise,target size,and rotation angle which interfere with feature extraction.Firstly,in the feature extraction network,a novel and efficient attention feature extraction module is introduced by combining an efficient layer aggregation network with a convolutional block attention module,enhancing both the quality and efficiency of feature extraction.Secondly,a fractional-order Gabor transform convolution module is constructed in the feature fusion network to emphasize fine-grained details such as the edges,textures,and orientations of aircraft targets,thereby improving feature fusion.Finally,a learnable dynamic detection head is applied in the detection layer,where a scale-aware attention module strengthens attention to multi-scale targets,a spatial-aware attention module enhances spatial position discrimination,and a task-aware attention module facilitates more precise distinction of task-specific requirements.Experimental results on the DOTAv1 dataset demonstrate that the proposed method achieves a detection accuracy of 96.2%,which is 2.2%higher than the baseline YOLOv7 model.The method also has a smaller model weight,with a notable improvement in detection accuracy in complex scenarios.This approach provides a more efficient solution for aircraft target detection in remote sensing images.

关 键 词：遥感图像 飞机目标检测 深度学习 卷积神经网络 分数阶Gabor变换 

分 类 号：V19[航空宇航科学与技术—人机与环境工程] TP751[自动化与计算机技术—检测技术与自动化装置]