融合多尺度特征的前景分割方法在高压开关柜异物遮挡检测中的应用

Foreground Segmentation with Multi-Scale Features for Foreign Object Detection of High Voltage Switchgear

电力系统巡检的一项任务是检查电力设备是否被异物遮挡,确保设备安全运行,并为元件状态巡检任务排除障碍。针对高压开关柜的异物遮挡检测任务,通过合成算法人工合成构建一个开关柜异物遮挡检测的合成数据集,然后将监控视频运动目标分割任务的前景分割方法应用于该任务,并通过融合多尺度特征提高模型对背景存在细微差别的分割精度。实验对比和分析几种主要的前景分割方法与所提方法的优劣结果证明针对任务改进的前景分割方法可以在开关柜异物遮挡检测任务中取得很好的效果。

A task of power systems inspection is to ensure whether a power equipment is occluded by foreign object so as to ensure the safety of the equipment and clear the barrier of instruments patrolling task. In this work,we construct a synthetic dataset for foreign object detection with the switchgear images for background and the foreign object images for foreground. Then,we manage to apply the method for moving object segmenting to our task,with using multi-scale features to improve the accuracy of segmenting the object in small distinct background. We compare the several foreground segmenting methods and our method used in the foreign object detection. The experiments show that our improved foreground segmenting method can achieve good results in the foreign object detection task of high voltage switchgear.