摘要
针对安全壳裂缝识别问题,引入迁移学习方法来提高卷积神经网络模型的训练效率。分析了各预训练卷积神经网络模型在基于迁移学习的安全壳裂缝识别中的效果。分别利用在ImageNet上预训练好的AlexNet、VGGNet、In⁃ception V3模型进行迁移学习,并用小样本数据集对这3种模型进行重新训练。结果表明,与重新训练的模型相比,迁移学习在减少了训练时间的同时还提升了分类任务的性能,其中,利用Inception V3预训练的权重参数进行迁移学习时表现最好,准确率可达97.16%。
To solve the problem of containment crack identification,transfer learning method is introduced to improve the training efficiency of convolutional neural network model.We analyze the effect of each pre-trained convolution neural network model in containment crack detection based on transfer learning.AlexNet,VGGNet and Inception V3 models pretrained on ImageNet are used for transfer learning,then small sample data sets are used to retrain these three models.The results show that,compared to the retraining model,transfer learning reduces training time and improves the performance of classification tasks.Among them,transfer learning performs best when using the weight parameters pre-trained by Inception V3,with an accuracy of 97.16%.
作者
徐亚明
虞剑
XU Yaming;YU Jian(School of Geodesy and Geomatics,Wuhan University,Wuhan 430079,China)
出处
《测绘地理信息》
CSCD
2023年第5期65-68,共4页
Journal of Geomatics
