基于改进YOLOX的变电站工人防护设备检测研究*

为解决电气工人防护设备检测问题，通过改进YOLOX算法，提出检测工作人员防护设备的模型。首先在预测部分改进损失函数，为解决损失函数计算存在的缺陷，对IOU损失的计算方法进行改进，根据防护设备任务特性，通过调整各种类型损失函数的权重，增加对模型误判的惩罚，对模型进行优化；其次在算法主干网络中引入CBAM注意力模块提高神经网络对工人防护设备的感知能力；最后在算法Neck部分，将UpSample结构用于多尺度特征融合，加强网络的细节表达能力，从而提升对小目标困难样本的检测精度。研究结果表明:改进后的YOLOX模型平均精度均值达到87.24%，与已有YOLOX模型相比提升2.46%，具备有效性，适用于变电站工人防护设备检测。研究结果可为电气工人提供更高的防护装备检测精度。

Research on detection of protection equipment for substation workers based on improved YOLOX

Aiming at the detection problem of electrical workers’ protective equipment,a model to detect the workers’ protective equipment was proposed by improving YOLOX (you only look once X) algorithm.Firstly,the loss function was improved in the prediction part,and the calculation method of intersection over union (IOU) loss was improved in order to solve the disadvantages of loss function calculation.At the same time,the model was optimized by adjusting the weights of various types of loss functions and increasing the penalties for model misjudgment according to the task characteristics of protective equipment.Secondly,the convolutional block attention module (CBAM) was introduced into algorithm backbone network to enhance the perception capability of neural network on workers’ protective equipment.Finally,the UpSample structure was introduced into the Neck part of the algorithm for multi-scale feature fusion to enhance the expression performance of network details,thus improve the recognition accuracy of small targets.The results showed that the mean average precision of the experiment on the data set reached up to as high as 87.24% by using the improved YOLOX model,which was 2.46% higher than the existing YOLOX model.It was effective and suitable for the detection of substation workers’ protective equipment.The research results can provide higher detection accuracy of protective equipment for electrical workers.