An Infrared Image Detection Method of Substation Equipment Combining Iresgroup Structure and CenterNet
Hanbo Zheng, Yaohui Cui, Wenqiang Yang, Jinheng Li, Liming Ji, Yuan Ping, Sijia Hu, Xin Chen
Abstract
The infrared image diagnostic technology of substation equipment is often interfered by its complex background, and cannot reflect the effective temperature information of the equipment, which is not conducive to the realization of subsequent fault diagnosis. In order to solve this problem, an improved CenterNet is proposed to improve the recognition effect of substation equipment under complex backgrounds. First, a richer substation infrared image dataset is established. Second, the Iresgroup structure is proposed as a backbone to further improve the feature learning ability of CenterNet. In the Iresgroup structure, an improved projection shortcut and Iresnet structure are improved to control the number of ReLUs on the main path, which can further improve the detection accuracy of the model, with higher robustness. In addition, the group conv structure is used to further improve the model, so that the model maintains the same amount of parameters while strengthening its feature learning ability. At last, The hyperparameters of centerNet have been redesigned to further improve the accuracy of the proposed method. The results show that compared with the several native backbones of the CenterNet and several mainstream detection models, this method has a significant improvement of 2% to 4% on mAP, and the improvement of single-type objects is as high as 7.4%. It can be seen that this method has achieved excellent detection results and can provide key technical support for fault detection of substation equipment.