基于DenseNet的唐卡图像分类  

Thangka image classification based on DenseNet

作　　者：曾富亮 胡文瑾[1,3] 何国源 薛盼盼 ZENG Fuliang;HU Wenjin;HE Guoyuan;XUE Panpan(School of Mathematics and Computer Science,Northwest Minzu University,Lanzhou 730030,China;Institute of China Ethnic Information,Northwest Minzu University,Lanzhou 730030,China;MOE Key Laboratory of National Languages Information Technology,Lanzhou 730030,China)

机构地区：[1]西北民族大学数学与计算机科学学院,甘肃兰州730030 [2]西北民族大学中国民族信息研究院,甘肃兰州730030 [3]教育部民族语言与信息技术重点实验室,甘肃兰州730030

出　　处：《现代电子技术》2022年第6期153-157,共5页Modern Electronics Technique

基　　金：国家自然科学基金项目:圣像唐卡图像的文本描述生成方法研究(62061042);国家自然科学基金项目:唐卡线描图生成及风格化创作关键技术研究(61862057);西北民族大学中央高校基本科研业务费资金资助项目(31920200066);国家民委创新团队计划资助(2018[98]号);西北民族大学“双一流”和特色发展引导专项资金资助项目。

摘　　要：针对DenseNet采用ReLU函数在特征传播过程中存在丢失图像的负特征问题,文中提出一种改进的DenseNet方法。首先,为了解决样本数据集少而产生的问题,采用数据增强方法中的图像翻转、旋转和随机裁剪方法来增加数据集;其次,利用Leaky ReLU函数不会丢失图像负特征的特性来代替ReLU函数,在反向传播时增加传播的负特征,并采用迁移学习的基于微调网络的技术进行训练从而防止过拟合。迁移学习是运用神经网络在大型数据集ImageNet上训练好的参数权重在小数据集上进行网络训练,再把原神经网络的全连接层改为小数据集需要的层。最后对唐卡图像进行分类。实验结果表明:改进的DenseNet取得了较好的性能,比DenseNet性能提高了1.1%;与VGG16、ResNet50和InceptionV3等其他卷积神经网络相比,改进的DenseNet对唐卡图像分类效果良好。In allusion to the problem that the ReLU function used by DenseNet can lose the negative features of the image during the feature propagation,an improved DenseNet method is proposed.The image flipping,rotation and random clipping method in the data enhancement method are used to increase the datasets to solve the problem caused by few sample datasets.The Leaky ReLU function,which has the feature that the negative features of the image cannot be lost,is used to replace the ReLU function.The negative features of the propagation are added during the back propagation,and the transfer learning technology based on the fine⁃tuning network is used for training,so as to prevent the over⁃fitting.The transfer learning is to use the parameter weights trained by neural network on ImageNet,a large dataset,to carry out network training on small datasets,then change the full connection layer of the original neural network to the layer required by small datasets,and finally classify Thangka images.The experimental results show that the improved DenseNet can achieve better performance,which is 1.1%higher than DenseNet,and has a good effect on Thangka image classification in comparison with other convolutional neural networks such as VGG16,ResNet50 and InceptionV3.

关 键 词：唐卡图像 图像分类 DenseNet 特征提取 数据集增强 网络训练 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391-41[电子电信—信息与通信工程]