基于改进Centernet的变电设备红外检测方法

Infrared detection method of substation equipment based on improved Centernet

变电站红外图像中小目标众多并且环境复杂,导致现有检测算法精度较低,因此本文提出一种基于改进Centernet的变电设备红外检测方法。首先以Centernet作为基础模型,将FPN结构引入上采样网络以充分利用小目标特征信息,从而解决小目标难以被精确检测的问题;然后,为提升网络在复杂环境中检测的鲁棒性,通过在主干网络resnet50中嵌入注意力机制来提升网络对重要目标的关注;最后,采用CIOU损失替换中心点偏移损失和宽高损失的训练策略以加速网络收敛、提升训练效果。实验结果表明,本文方法在小目标检测和复杂环境检测中都能有较好的检测效果,检测精度相比改进前提升3.1%,达到92.7%,相比Faster R-CNN等现有方法精度更高,在变电设备红外检测中具有一定参考价值。

There are many small targets in infrared images of substations with complex environment,resulting in low accuracy of existing detection algorithms.Therefore,this paper proposes an infrared image detection method for substation equipment based on improved Centernet.Firstly,taking Centernet as basic model,the FPN structure is introduced into the upsampling network to fully use the feature information of small targets,so as to solve the problem that small targets are difficult to be accurately detected;Then,in order to improve the detection robustness of the network in complex environment,an attention mechanism is embedded in the backbone network resnet50to increase the attention of network to core targets;Finally,the training strategy of center point offset loss and width and height loss is replaced by CIOU loss to accelerate network convergence and improve training effect.The experimental results show that the method in this paper can have better detection effect in both small targets detection and complex environment detection,and the detection accuracy is improved by 3.1%,reaching 92.7%,which is more accurate than existing methods such as Faster R-CNN,and has certain reference value in infrared image detection of substation equipment.