融合空间信息的多尺度残差卷积开关柜异物检测算法研究

Research on Foreign Object Detection Algorithm of Multi-scale Residual Convolutional Switchgear Fused With Spatial Information

检查开关柜中有无异物以确保设备安全运行是配网建设的基本任务。针对开关柜中背景复杂且目标遮挡严重的问题,文章提出一种融合空间信息的多尺度残差卷积检测算法。首先,利用多尺度残差卷积,降低目标遮挡导致特征提取不全的影响,再设置残差连接解决过拟合问题;然后,在深层特征图间增加改进的注意力机制,降低网络过深导致细节信息丢失影响,提高检测效果。最后,搭建自制开关柜异物数据集实验平台,实验结果表明,改进模型检测速度下降11FPS(frames per second),为72FPS,平均精度AP50为91.26%,AP@50:5:95为76.04%,分别提高2.59%和3.69%。并在输电线路异物与缺陷绝缘子数据集验证普适性算法,实验结果表明,改进模型检测精度均高于原始模型。

Checking whether there are foreign objects in the new switch cabinet to ensure the safe operation of equipment is the basic task of distribution network construction.Aiming at the problem of complex background and serious target occlusion in the switch cabinet,a multi-scale residual convolution detection algorithm that integrates spatial information is proposed.Firstly,multi-scale residual convolution is used to reduce the impact of incomplete feature extraction caused by target occlusion,and then set the residual connection to solve the overfitting problem.Finally,the improved attention mechanism for channel and space integration is added between the deep feature maps,which reduces the effect of the loss of small target features caused by the network too deep,and improves the detection effect of small target objects.Eventually,a dataset of the switch cabinet for foreign-object detection experiment platform is made.In the experiment on the foreign body dataset of the self-made switchgear,the detection speed of the improved model decreased by 11FPS(Frames Per Second)to 72FPS,and the average accuracy AP50 was 91.26%,which was 76.04%compared with the AP@50:5:95,which was increased by 2.59%and 3.69%,respectively.Experiments have proved that the detection method has a high detection accuracy and running speed,and has practical application value.