摘要
针对壁画数目较少会导致深度学习模型训练不够充分的问题,提出了一种结合深层迁移模型和底层简单视觉相结合的古代壁画分类模型。模型将Visual Geometry Group Network迁移模型学习到的知识应用到目标小样本集上进行微调得到壁画高层特征,结合颜色直方图和局部二值模式纹理算子为特征描述子进行分类。实验结果表明:在构建的壁画图像数据集上,模型达到了90.35%的准确率。与传统模型和其他的卷积神经网络相比,算法在分类准确率、泛化能力和稳定性上都有很大提升,验证了模型对于壁画分类的有效性。
Aiming at the problem that a small number of murals will lead to insufficient training of deep learning model,an ancient mural classification model combining deep migration model with low-level vision is proposed.The model applies the knowledge learned from the Visual Geometry Group Network migration model to the small sample set of the target for fine-tuning to obtain the high-level characteristics of murals and combines color histogram and local binary patterns texture histogram as feature descriptors for classification.The experimental results show that the model achieves an accuracy of 90.35%on the mural image data set.Compared with the traditional model and other convolutional neural networks,the algorithm improves the classification accuracy,generalization ability and stability.The effectiveness of the mural classification model has been proved.
作者
曹建芳
崔红艳
张自邦
赵爱迪
Cao Jianfang;Cui Hongyan;Zhang Zibang;Zhao Aidi(College of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030000,China;Department of Computer,Xinzhou Teachers University,Xinzhou 034000,China)
出处
《系统仿真学报》
CAS
CSCD
北大核心
2021年第5期1095-1103,共9页
Journal of System Simulation
基金
山西省自然科学基金(201701D21059)
山西省高等学校人文社会科学重点研究基地项目(20190130)
山西省艺术科学规划课题(2017F06)
忻州市平台和人才专项(20180601)。
关键词
迁移学习
壁画分类
特征融合
底层特征
transfer learning
mural classification
feature fusion
low-level features
