| 摘要: |
| 目的 绝缘子缺陷的定期检测与维修对保障输电线路的安全有至关重要的作用,为了解决绝缘子缺陷检测
方法存在检测精度不高、泛用性不强等问题,提出了一种基于改进 DETR( Detection Transformer) 的绝缘子缺陷检测
算法。 方法 设计改进编码器,使用 4 个 Transformer stage 来捕捉图像中不同尺度和关系的特征信息。 同时,还利用
了 ResNet50 的中间输出特征来补充分层 Transformer 的输出特征,从而提升目标检测算法的性能。 设计改进解码
器,采用了三层串联的结构,以确保解码器能够在不同阶段接收并学习不同尺度的特征图,同时特征融合增强模块
和查询更新模块使解码器能够更有效地学习图像的特征信息且降低匹配具有相似语义特征区域的难度,进一步提
高网络检测的准确率。 结果 通过对输电线路绝缘子缺陷航拍图像进行了仿真实验研究,在不同阈值下改进方法识
别精度分别达到了 99. 5%、80. 4%,较原算法分别提升了 3. 4%、6. 1%,对部分遮挡目标有较好的检测效果,同时与
其他算法相比具有更优的检测精度和泛化能力。 结论 改进 DETR 具有更高的检测性能,实现对绝缘子缺陷的准确
检测,为下一步对于其他输电线路目标,如防震锤、间隔棒等检测提供了保证。 |
| 关键词: 绝缘子缺陷 改进 DETR 空洞卷积 DIOU |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Insulator Defect Detection Algorithm Based on Improved DETR |
|
OUYANG Mingsan LI Jie
|
|
School of Electrical and Information Engineering Anhui University of Science and Technology Anhui Huainan 232000
China
|
| Abstract: |
| Objective Regular inspection and maintenance of insulator defects play a crucial role in ensuring the safety of
transmission lines. In order to address issues such as low detection accuracy and poor universality of existing insulator
defect detection methods an algorithm based on an improved Detection Transformer DETR was proposed. Methods
An improved encoder was designed to use four Transformer stages to capture the feature information of different scales and
relationships in the image. At the same time the intermediate output features of ResNet50 were also used to supplement
the output features of the layered Transformer thereby enhancing the performance of the object detection algorithm. The
improved decoder was designed and a three-layer series structure was adopted to ensure that the decoder can receive and
learn feature maps of different scales at different stages. Moreover the feature fusion enhancement module and the query
update module can make the decoder learn image feature information more effectively and reduce the difficulty of matching
regions with similar semantic features further improving the accuracy of network detection. Results Simulation
experiments were conducted on aerial images of insulator defects in transmission lines. The improved method achieved
recognition accuracies of 99. 5% and 80. 4% at different thresholds respectively which were 3. 4% and 6. 1% higher
than those of the original algorithm. It exhibited good detection performance for partially occluded targets and demonstrated superior detection accuracy and generalization ability compared with other algorithms. Conclusion The improved DETR
demonstrates higher detection performance enabling accurate detection of insulator defects. This provides assurance for
the detection of other targets on transmission lines in the future such as vibration dampers and spacer rods. |
| Key words: insulator defect improved DETR dilated convolution DIOU |
