| 禹鑫燚,林密,卢江平,欧林林.基于向量叉乘标签分配的遥感图像目标检测算法[J].高技术通讯(中文),2024,34(2):132~142 |
| 基于向量叉乘标签分配的遥感图像目标检测算法 |
| Remote sensing object detection algorithm based on vector cross product label assignment |
| |
| DOI：10. 3772/ j. issn. 1002-0470. 2024. 02. 003 |
| 中文关键词: 遥感图像； 目标检测； 标签分配； 向量叉乘 |
| 英文关键词: remote sensing image, object detection, label assignment, vector cross product |
| 基金项目: |
| 作者 | 单位 | | 禹鑫燚 | （浙江工业大学信息工程学院杭州 310023） | | 林密 | | | 卢江平 | | | 欧林林 | |
|
| 摘要点击次数: 454 |
| 全文下载次数: 303 |
| 中文摘要: |
| 近年来遥感图像目标检测受到了广泛的关注，主流的遥感图像目标检测器通过预设锚框与真实框之间的交并比（IoU）进行正负样本的划分。为了解决基于IoU的标签分配方法在遥感图像小而密集目标中存在复检和漏检的问题，本文提出了一种基于向量叉乘标签分配的遥感图像目标检测算法YOLOXR。首先，提出了一种标签粗分配策略，通过向量叉乘的方法判断特征图的像素点是否在旋转目标内或者目标中心点附近的旋转正方形框内，从而确定其是否为候选正样本。其次，为了降低边缘低质量候选正样本对标签分配的影响，提出了旋转中心度量方法，通过向量叉乘判断像素点距离中心点的远近程度进而赋予不同的权重。最后，基于最优传输的方法(simOTA)选取真实框和样本点的最优匹配对，使得总体代价最小，进而为旋转目标分配合适的标签。此外，为了解决旋转IoU损失不可导以及Smooth L1损失难以权衡旋转框各个参数的问题，通过计算真实框和预测框二维高斯分布的Kullback-Leibler散度(KLD)来替代IoU。在公开的遥感图像目标检测数据DOTA、HRSC 2016和UCAS-AOD上的大量实验表明，所提方法优于目前绝大多数旋转目标检测算法。 |
| 英文摘要: |
| Object detection in aerial images has received extensive attention in recent years, and the mainstream remote sensing image object detectors divide positive and negative samples by the intersection-over-union (IoU) between the preset anchor box and ground-truth box. In order to solve the problem of duplicate detection and missed detection in remote sensing images with small and dense objects in the label assignment method based on IoU, a remote sensing image object detection method YOLOXR based on vector cross product label assignment is proposed. Firstly, a rough label assignment strategy is proposed, which uses the vector cross product method to determine whether a pixel is in the oriented object or the rotating square box near the center of the object, so as to determine whether it is a candidate positive sample. Secondly, a rotation center measurement approach is provided to limit the influence of low-quality candidate positive samples on label assignment by judging the distance between the pixel point and the center point using vector cross product and then assigning different weights. Finally, optimal transmission assignment (simOTA) is used to select the optimal matching pairs of ground-truth boxes and the sample points, which minimizes the overall cost, and then assigns the appropriate label to the rotating object. In addition, IoU is replaced by computing the Kullback-Leibler divergence (KLD) of the two-dimensional Gaussian distribution of the ground-truth box and the predicted box to overcome the problem that the rotation IoU loss is not differentiable and the Smooth L1 loss is difficult to be used to weigh the parameters of the oriented bounding boxes. Extensive experiments on public remote sensing image object detection datasets DOTA, HRSC 2016 and UCAS-AOD show that the proposed method outperforms most of the current oriented object detection algorithms. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
