基于改进RetinaNet的电力设备红外目标精细化检测模型  

Refined Infrared Object Detection Model for Power Equipment Based on Improved RetinaNet

作　　者：苏海锋[1] 赵岩[1] 武泽君 程博[1] 吕林飞 SU Haifeng;ZHAO Yan;WU Zejun;CHENG Bo;LYU Linfei(Department of Electrical Engineering,North China Electric Power University,Baoding 071000,China)

机构地区：[1]华北电力大学电力工程系,河北保定071000

出　　处：《红外技术》2021年第11期1104-1111,共8页Infrared Technology

基　　金：国家重点研发计划(2017BJ0080)。

摘　　要：电力设备在运行过程中会产生大量红外图像,当红外图像中的电力设备存在排列密集、具有倾斜角度、大长宽比的情况时,基于水平矩形框的目标检测网络只能给出目标概略位置,易发生目标检测区域重叠,引入冗余背景信息,使得检测结果不够精细。针对此问题,提出在RetinaNet目标检测网络中引入旋转矩形框机制,并在网络输入端引入Mosaic数据增强技术;将原特征提取网络中ReLU函数替换为梯度流更平滑的Mish激活函数;在原模型FPN模块后追加PAN模块进一步融合图像特征。最后利用现场采集的电力设备红外图像制作数据集,将改进后的模型与Faster R-CNN、YOLOv3、原RetinaNet三种基于水平矩形框定位的目标检测网络进行对比评估,实验表明改进后的模型可以更为精细地检测出密集场景下带有倾角的电力设备红外目标,在多类别电力设备检测准确率对比上高于以上3种模型。A large number of infrared images are generated during the operation of power equipment.When the power equipment in the infrared image is densely arranged,incline-angled,and has a large aspect ratio,the target detection network based on a horizontal rectangular frame can only provide the approximate posi-tion of the target,which is prone to overlap with the target detection area and introduce redundant back-ground information,giving detection results that are not sufficiently accurate.To solve this problem,we propose to introduce a rotating rectangular box mechanism into the retina net target detection network and mosaic data enhancement technology at the network input,replacing the ReLU function in the original backbone network with a smoother mish activation function of gradient flow;the Pan module is added after the FPN module of the original model to further fuse image features.Finally,the data set is made by using the power-equipment infrared images collected on-site.The improved model is compared and evaluated with three target detection networks based on horizontal rectangular frame positioning:fast R-CNN,YOLOv3,and original RetinaNet.The experiments show that the improved model can detect the infrared targets of power equipment with inclination in dense scenes more accurately,and the detection accuracy of mul-ti-category power equipment is higher than that of the above three models.

关 键 词：红外图像 RetinaNet 电力设备 卷积神经网络 目标检测 

分 类 号：TM85[电气工程—高电压与绝缘技术]