摘要
由于变电站背景下对于不同仪表而言相似度极高,且存在光照、噪声、磁场等影响因素,复杂情况下的分类识别率并不高,传统视觉特征的分类准确率不能满足要求,误差较大。针对特定指针仪表目标的图像分类问题,通过利用AlexNet卷积神经网络构,将构造的特定背景下仪表数据集导入网络之中进行训练。训练得到验证集精度达到94.76%。然后从网络结构中池化方式的不同这方面进行改进,得到一种新的组合方式,既保留了轮廓的完整性,又在细节的处理上更加精确。有效保留了数据间的相关联性,间接加强了模型的学习能力,提高了目标分类的准确率,使得验证集精度提高到97.68%。
Due to the high similarity of different instruments under the background of substation,and the influence factors of illumination,noise and magnetic field,the classification recognition rate is not high under complex conditions,and the classification accuracy of traditional visual features can not meet the requirements,and the error is large.Aiming at the image classification problem of specific pointer instrument target,the instrument data set under specific background is imported into the network for training by using AlexNet convolution neural network.The accuracy of the verification set is 94.76%.Then,a new combination method is obtained by improving the different pooling methods in the network structure,which not only retains the integrity of the contour,but also is more accurate in the details processing.It can effectively preserve the correlation between data,indirectly enhance the learning ability of the model,improve the accuracy of target classification,and make the accuracy of verification set increase to 97.68%.
作者
胡鑫
欧阳华
尹洋
HU Xin;OUYANG Hua;YIN Yang(College of Electrical Engineering,Naval University of Engineering,Wuhan 430033)
出处
《舰船电子工程》
2021年第10期112-116,共5页
Ship Electronic Engineering
