基于改进YOLOv7的遥感图像旋转目标检测  

Rotating object detection in remote sensing image based on improved YOLOv7

作　　者：崔家礼[1] 刘远 CUI Jiali;LIU Yuan(School of Information Science and Technology,North China University of Technology,Beijing 100144,China)

机构地区：[1]北方工业大学信息学院,北京100144

出　　处：《微电子学与计算机》2025年第4期48-57,共10页Microelectronics & Computer

基　　金：北航杭州创新研究院钱江实验室开放基金(2020-Y3-A-014)。

摘　　要：遥感图像目标的高效精确检测是目标检测领域的重要问题。然而,物体有限的外观纹理特征和多样的旋转方向使得遥感图像目标检测变得困难。针对这些问题,提出了一种改进YOLOv7的遥感图像旋转目标检测算法。首先,引入KL(Kullback-Leibler)散度作为回归损失函数将旋转框坐标转换为二维高斯分布,解决了传统水平框检测在计算旋转角度时产生边界不连续的问题。其次,引入选择性大核卷积改造YOLOv7网络的特征提取模块,增强网络对目标形状、类别、尺度等特征信息的感知能力,提高网络模型的精度。最后,针对检测头中分类和回归任务共享特征带来的精度下降问题,采用了TSCODE特征解耦的检测头,提升了网络对分类特征和回归特征的学习能力。在DOTAv1.0和HRSC2016数据集上进行了相关实验,验证了所提方法的有效性和鲁棒性。Efficient and accurate detection of targets in remotely sensed images is an important problem in the field of target detection.However,the limited appearance texture features of objects and their different rotation directions make target detection in remote sensing images challenging.To address these issues,this paper proposes an improved YOLOv7-based algorithm for detecting rotated targets in remote sensing images.Firstly,Kullback-Leibler(KL)divergence is introduced as a regression loss function to transform the coordinates of rotated bounding boxes into a two-dimensional Gaussian distribution,which solves the problem of discontinuous boundaries in traditional horizontal box detection when calculating rotation angles.Secondly,selective large kernel convolution is introduced to modify the feature extraction module of the YOLOv7 network,which enhances the network's ability to perceive target shape,category,scale and other feature information,thus improving the accuracy of the network model.Finally,to address the accuracy degradation caused by the sharing of features between classification and regression tasks in the detection head,a TSCODE feature decoupling detection head is employed to enhance the network's learning capability for classification and regression features.The experiments conducted on the DOTAv1.0 and HRSC2016 datasets validate the effectiveness and robustness of the proposed method.

关 键 词：遥感图像旋转检测 密集场景 选择性大核卷积 渐进式融合解耦检测头 YOLOv7 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]