基于SGD和余弦退火算法改进YOLOv3的高压电力设备目标检测方法

An Improved YOLOv3 Target Detection Method for High Voltage Power Equipment Based on SGD and Cosine Annealing Algorithm

针对现有高压电力设备检测方法存在实时性差、准确性低和难以部署在移动端等问题,提出一种基于随机梯度下降(SGD)和余弦退火算法改进YOLOv3的高压电力输送设备安全检测算法。采用网络复杂度较小、计算速度快、识别精度高且易于部署的移动端YOLOv3作为算法的主要框架;然后设计了深层的残差网络(Darknet53)作为该模型的主干特征提取网络,在提高识别精度的同时解决网络过深可能产生的梯度爆炸问题;进一步地结合SGD优化算法和余弦退火算法,在保证网络训练学习效率较高的同时避免网络陷入局部最优解,以此提高高压电力设备安全检测的速度和精度,满足实际需要;最后使用采集的高压电力设备数据集对整个网络进行训练。结果表明,YOLOv3在高压电力设备数据集上的平均检测精度达到了97.08%,检测速度达到了56帧/s,误检率只有0.78%。

Aiming at the problems of poor real-time performance,low accuracy and difficult deployment in mobile terminal of the existing high-voltage power equipment detection methods,stochastic gradient descent(SGD)and cosine annealing algorithms are proposed to improve the YOLOv3 security detection algorithm for high-voltage power transmission equipment.Darknet53 was used as the feature extraction network of the model,which improves the detection speed of the traditional detection method of high voltage power equipment.Then,the SGD optimization algorithm and cosine annealing algorithm are used to improve the safety detection accuracy of high-voltage power equipment and accelerate the convergence speed of the model in the early stage.Finally,the collected high-voltage power equipment data set is used to train the entire network is trained.The results show that the average detection accuracy of the YOLOv3 model on the high-voltage power equipment dataset reaches 97.08%,the detection speed reaches 56 frames/s,and the false detection rate is only 0.78%.