基于多尺度特征融合的绝缘子缺陷程度检测

Insulator Defect Degree Detection Based on Multi-scale Feature Fusion

针对绝缘子不同程度缺陷特征相似、像素信息少、不同程度缺陷检测效果不佳的问题,提出了一种基于多尺度特征融合的绝缘子缺陷程度检测网络(multi-scale feature fusion defect degree detection network,MFFD3Net)。该网络采用重构的ResNeSt50架构提高了对绝缘子缺陷程度数据集的特征提取能力。设计了基于反卷积的多尺度特征融合模块,丰富了不同尺寸特征图的表达能力,提高了对不同尺度目标的检测性能。同时,在输入检测模块的浅层特征图后增加多感受野的特征提取模块(receptive field block,RFB),使得更多绝缘子缺陷信息进入有效感受野,对最终特征图产生影响,提升不同程度绝缘子缺陷的检测精度。MFFD3Net在绝缘子缺陷程度数据集上的全类平均精度达到85.02%,其中绝缘子轻微破损与绝缘子轻微闪络小目标的检测精度分别为78.37%、79.98%,能够完成不同程度绝缘子缺陷的识别与定位。因此,该文提出的MFFD3Net对于完善电力系统故障预警、保障电网安全稳定运行具有重要意义。

Insulator defects with different degrees have similar features and less pixel information,resulting in poor detection effect,therefore,an insulator defect degree detection network based on multi-scale feature fusion(MFFD3Net)is proposed.The network uses reconstructed ResNeSt50 to improve the feature extraction ability in insulator defect dataset.A multi-scale feature fusion module based on deconvolution is designed,which enriches the expression ability of different size feature maps and improves the detection performance of different scale targets.At the same time,the receptive field block(RFB)is added after the shallow feature maps of the input detection module to ensure more insulator defect information to enter the effective receptive field,which has an impact on the final feature map and improves the detection accuracy of insulator defects in different degrees.The mAP of MFFD3Net on insulator defect degree dataset reaches 85.02%,the detection accuracy of small targets such as slight breakage and slight flashover is 78.37%and 79.98%,which can complete the identification and location of insulator defects in different degrees.Thus,the MFFD3Net proposed in this paper is of great significance for improving the fault warning of power system and ensuring the safe and stable operation of power grid.