| 基于水平框弱监督标注的绝缘子旋转目标检测 |
投稿时间：2026-05-18 修订日期：2026-06-30  点此下载全文 |
| 引用本文： |
| 摘要点击次数: 22 |
| 全文下载次数: 0 |
|
| 基金项目:贵州电网有限责任公司都匀供电局科技项目(GZKJXM20240539) |
|
| 中文摘要:针对现有水平框监督旋转目标检测方法在角度预测稳定性、形状一致性方面存在的不足，本文提出了一种改进的 H2R-IWSD（H2RBox based Insulator Weakly Supervised horizontal Detection）检测算法。该算法在 H2RBox-v2 的基础上，首先引入了对数空间形状一致性损失，通过约束不同视图下目标宽高的对数变换结果相等，解决了原始方法仅关注角度一致性而忽略形状信息的问题，并设计了线性预热机制避免训练初期损失震荡。其次，对单位循环角度编码器进行了全面升级，增加了可学习基函数和轻量级注意力机制以提升角度表示能力，引入Von Mises 周期性损失更好地建模角度的周期性特性，同时提出了基于宽高比与分类置信度的动态阈值过滤策略，有效减少了无效角度预测。在IDID和CPLID数据集的实验结果表明，H2R-IWSD 相比RSAR在 mAP 指标上提升了 0.96%和1.33%，尤其在细长目标检测任务中表现出显著优势，同时保持了与原始方法相当的推理速度。 |
| 中文关键词:弱监督目标检测 有向目标检测 水平框监督 角度编码 |
| |
| Adaptive Rotation Detection for Insulators Based on Weakly Supervised Horizontal Bounding Box Annotations |
|
|
| Abstract:Existing horizontally supervised rotated object detection methods suffer from deficiencies in angle prediction stability and shape consistency. To address this issue, this paper presents the Horizontal-to-Rotated Improved Weakly Supervised Detector (H2R-IWSD). First, a log-space shape consistency loss is introduced. It constrains the logarithmic transformation of object width and height under different views to maintain consistency, remedying the defect of original methods that only focus on angle consistency while neglecting shape information. Meanwhile, a linear warm-up mechanism is devised to avoid loss oscillation in the early training stage. Second, the unit quaternion-based angle encoder is comprehensively upgraded. Learnable basis functions and a lightweight attention mechanism are embedded to strengthen angle representation capability, and the von Mises periodic loss is adopted to characterize the periodic property of rotation angles. Furthermore, a dynamic threshold filtering strategy based on aspect ratio and classification confidence is presented to effectively eliminate invalid angle predictions. Experimental evaluations are conducted on the IDID, CPLID, and remote sensing insulator datasets. The results demonstrate that the proposed H2R-IWSD surpasses the RSAR by 0.96% and 1.33% in mAP. It exhibits distinct superiority in detecting slender targets and preserves a comparable inference speed with the baseline method. |
| keywords:weakly supervised object detection oriented object detection horizontal bounding box supervision angle encoding |
| 查看全文 查看/发表评论 下载pdf阅读器 |
位访问者,今日访问
次