| 禹鑫燚,卢江平,林密,周利波,欧林林.基于单阶段全卷积检测器的遥感图像形状自适应椭圆标签分配方法[J].高技术通讯(中文),2024,34(8):875~884 |
| 基于单阶段全卷积检测器的遥感图像形状自适应椭圆标签分配方法 |
| Shape-adaptive ellipse label assignment for remote sensing image based on FCOS |
| |
| DOI：10. 3772 / j. issn. 1002-0470. 2024. 08. 009 |
| 中文关键词: 遥感图像； 深度学习； 目标检测； 标签分配 |
| 英文关键词: remote sensing image, deep learning, object detection, label assignment |
| 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） | | 卢江平 | | | 林密 | | | 周利波 | | | 欧林林 | |
|
| 摘要点击次数: 449 |
| 全文下载次数: 605 |
| 中文摘要: |
| 基于无锚框的检测方法在目标检测领域中发展迅速。然而在遥感图像中，目标存在角度任意、密集排列以及形状差异大等难点，使得遥感图像的检测仍是一项挑战。为此，本文提出了基于单阶段全卷积检测器(FCOS)改进的无锚框检测方法。首先，为了挖掘更多潜在的高质量锚点，提出基于椭圆方程的形状自适应特征点采样方法。然后，为进一步降低边界低质量样本点的影响，提出椭圆中心度量方法，相较原有的中心度量方法提供更合理的权重。此外，针对分类与回归的不一致问题，提出交并比（IoU）联合指导策略，将椭圆中心度量与IoU得分相结合作为质量分数监督分类分支，进一步提升检测精度。在DOTA 1.0数据集上的平均精度达到了79.17%，优于现有多数无锚框检测算法。 |
| 英文摘要: |
| Anchor-free object detection algorithms have experienced rapid development in object detection in recent years. However,in remote sensing images, the objects with arbitrary angles, dense distribution, and large shape differences make the detection still a challenge. Therefore, an anchor-free method based on improved fully convolutional one-stage (FCOS) is proposed. Firstly, to mine more potential high-quality anchor points, a shape-adaptive feature point sampling method based on the ellipse equation is proposed. To further reduce the negative influence of low-quality anchor points, the ellipse centerness is proposed. It can provide more accurate and reasonable weights than the traditional centerness. In addition, to address the inconsistency between classification and regression, a joint intersection over union (IoU) guidance strategy is proposed. The proposed ellipse centerness and IoU score are combined as quality scores to guide the training of the classification branch and to make the results of regression more accurate. The mean average precision on the DOTA 1.0 dataset reaches 79.17%, which is better than most existing anchor-free detection methods. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
