[1]陈玉红,刘晓静*.基于卷积神经网络的唐卡尊像自动分类研究[J].计算机技术与发展,2021,31(12):167-174.[doi:10. 3969 / j. issn. 1673-629X. 2021. 12. 028]
　CHEN Yu-hong,LIU Xiao-jing*.Research on Automatic Classification of Thangka Images Based on Convolutional Neural Network[J].,2021,31(12):167-174.[doi:10. 3969 / j. issn. 1673-629X. 2021. 12. 028]
点击复制

基于卷积神经网络的唐卡尊像自动分类研究()

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
31
期数:
2021年12期
页码:
167-174
栏目:
应用前沿与综合
出版日期:
2021-12-10

文章信息/Info

Title:
Research on Automatic Classification of Thangka Images Based on Convolutional Neural Network
文章编号:
1673-629X(2021)12-0167-08
作者:
陈玉红; 刘晓静*
青海大学 计算机技术与应用系,青海 西宁 811601
Author(s):
CHEN Yu-hong; LIU Xiao-jing*
Department of Computer Technology and Application,Qinghai University,Xining 811601,China
关键词:
非物质文化遗产; 数字化保护; 卷积神经网络; 图像分类; 唐卡尊像数据集
Keywords:
intangible cultural heritage; digital protection; convolutional neural network; image classification; dataset of Thangka images
分类号:
TP181
DOI:
10. 3969 / j. issn. 1673-629X. 2021. 12. 028
摘要:
唐卡作为中国的非物质文化遗产之一,受到越来越多人的关注。 如何快速、准确地对唐卡中的尊像进行分类,对于唐卡的研究以及数字化保护传承极其重要。 因此,该文提出了一种改进的基于卷积神经网络基本结构的唐卡尊像自动分类方法。 通过手动采集及爬虫技术等收集唐卡图像,构建唐卡尊像数据集,并作为网络模型的输入数据。 在保留原有卷积神经网络基本结构的前提下,在传统卷积神经网络结构每组的隐藏层中加入批量归一化层,改善模型的训练效率,并以 Relu 作为卷积池化层的激活函数,在最后一层全连接输出层前面加入 Dropout 层,减少过拟合,全连接输出层使用Softmax 作为激活函数,而损失函数则使用交叉熵,使分类效果更好,同时采用 Adam 优化方法来进行模型的优化,并应用在唐卡尊像分类方面。 最终在自己建立的唐卡数据集上进行实验,分类准确率高达 94. 7% ,比其他典型方法高出约 3% ,分类效果更佳,更有利于唐卡这种非物质文化遗产的数字化保护。
Abstract:
As one of China’s intangible cultural heritage,Thangka has attracted more and more attention. How to quickly and accurately classify Thangka images is extremely important for the research and digital protection of Thangka. Therefore,we propose an improved automatic classification method for Thangka images based on the basic structure of the convolutional neural network. Thangka images are collected by manual collection and crawler technology, and the Thangka image data set is constructed and used as the input data of network model. Under the premise of retaining the basic structure of the original convolutional neural network, each group of the traditional convolutional neural network structure adds a batch normalization layer to the hidden layer to improve the training efficiency of the model and uses Relu as the activation function. A Dropout layer is added before the last fully connected output layer to reduce over fitting and fully connected output layer uses Softmax as the activation function,and the loss function uses cross-entropy to make the classification effect better. At the same time,the Adam optimization method is used to optimize the model and applied to the classification of Thangka images. Finally,the experiments on the Thangka data set established by ourselves show that the classification accuracy rate is as high as 94. 7% ,which is about 3% higher than other typical methods. The classification effect is better,which is more conducive tothe digital protection of Thangka,an intangible cultural heritage.

相似文献/References:

[1]陈世婕,王卫星*,彭 莉.基于多尺度网络的苗绣绣片纹样分割算法研究[J].计算机技术与发展,2023,33(11):149.[doi:10. 3969 / j. issn. 1673-629X. 2023. 11. 022]
　CHEN Shi-jie,WANG Wei-xing*,PENG Li.Research on Miao Embroidery Pattern Segmentation Algorithm Based on Multi-scale Network[J].,2023,33(12):149.[doi:10. 3969 / j. issn. 1673-629X. 2023. 11. 022]

更新日期/Last Update: 2021-12-10