摘要
针对人工处理重复供电安全管控图像效率低的问题,在感知哈希算法(perceptual Hash algorithm,PHA)的基础上,给出了基于分块离散余弦变换(block discrete cosine transform,BDCT)的PHA,简称BDCT-PHA。采用BDCT-PHA进行图像去重操作,该算法可对经过JPEG压缩的图像进行处理,具有较高的去重准确率和较低的误判率。然后,改进ResNet网络结构并选择卷积神经网络(convolutional neural network,CNN)进行图像分类操作,先将图像转换为VOC数据集的格式,再对其训练,该方法防止了网络模型加深时出现的梯度消失现象,在减少计算量的同时提高了分类效率。仿真实验表明:提出的方法能够准确识别出重复图片,并标明相似的图像的编号,在对供电安全管控图像分类时,能够使损失函数值收敛在4.785%,分类准确率高达94.46%。
In view of the low efficiency of manual processing of repeated power supply security control image,based on the perceptual Hash algorithm(PHA),PHA based on block discrete cosine transform hash(BDCT)was proposed,abbreviated as BDCT-PHA.BDCT-PHA was used for image duplication removal.The algorithm could process the JPEG compressed image,and had high duplication removal accuracy and low misjudgment rate.Then,the ResNet network structure was improved,and the convolutional neural network(CNN)was selected for image classification.Firstly,the image was transformed into the format of VOC data set,and then trained.This method prevented the gradient disappearance when the network model was deepened,and improved the classification efficiency while reducing the amount of calculation.Simulation results show that the proposed method can accurately identify duplicate images and indicate the number of similar images.When classifying power supply safety management and control images,the loss function value can converge to 4.785%and the classification accuracy is as high as 94.46%.
作者
曹增新
蒋程
朱龙辉
CAO Zengxin;JIANG Cheng;ZHU Longhui(State Grid Beijing Fangshan Power Supply Company,Beijing 102401,China;School of Electrical Engineering,Xi’an University of Technology,Xi’an 710054,China)
出处
《西安工程大学学报》
CAS
2021年第6期62-68,75,共8页
Journal of Xi’an Polytechnic University
基金
国家重点研发计划(2019YFE0123600)。
