基于多时相巡检图像的变电设备抗干扰缺陷检测

近年来,变电站中广泛采用机器视觉算法分析多时相巡检图像的差异变化,用于检测各类变电设备缺陷,以确保运行安全.然而,由于拍摄时刻不同,多时相图像间存在天气、光照、季节等各类干扰变化,对变电设备的缺陷检测提出了挑战.对此,提出一种基于多时相巡检图像的变电设备抗干扰缺陷检测方法.首先,利用风格迁移模型CycleGAN学习不同风格域之间的映射关系,并基于检测图生成足量存在天气、光照、季节干扰变化的干扰图;其次,基于参考图$+$检测图$+$干扰图三元组对三重孪生网络TripleNet进行协同训练,在特征层面提出空间一致性损失以抵抗各类干扰变化,用于提取三者鲁棒的多尺度差异特征;最后,搭建特征聚合网络PANet融合多尺度差异特征,输出多尺度的缺陷检测结果.在实际变电设备多时相巡检图像数据集上进行实验验证,结果表明,所提出方法相较于非孪生网络和一般孪生网络可提升2.09%和0.67%的mAP,且在原始样本与干扰样本上的检测精度更均衡,而且所提出方法可以在提升变电设备缺陷检测模型精度的同时增强模型的抗干扰能力.

Anti-interference defect detection of substation equipment based on multi-temporal inspection images

In recent years, machine vision algorithms are widely used in substations to analyze the difference changes of multi-temporal inspection images, which are used to detect various substation equipment defects to ensure the safety of operation. However, due to the different shooting times, there are various interference changes such as weather, illumination and season between multi-temporal images, which pose challenges to the defect detection of substation equipment. Therefore, this paper presents an anti-interference defect detection method for substation equipment based on multi-temporal inspection images. Firstly, the style transfer model CycleGAN is utilized to learn the mapping between different style domains, then interference images with weather, light and seasonal interference changes are generated from the detected images. Secondly, we utilize the reference image, the detection image and the interference image to train a TripleNet cooperatively, and a spatial consistency loss is proposed to resist various interference changes at the feature level, which aims to extract the robust multi-scale difference features. Finally, a path aggregation network is built to fuse multi-scale difference features, which is utilized to get multi-scale defect detection results. The experimental verification is carried out on the multi-temporal inspection image dataset of actual substation equipment. Compared with the non siamese network and the general siamese network, the proposed method can improve the mAP by 2.09% and 0.67%, and the accuracy of original samples and interference samples is more balanced. The experiments demenstrate that the proposed method can improve the accuracy and enhance the anti-interference ability of the model for substation equipment defect detection.