欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2021, Vol. 42 ›› Issue (5): 744-754.DOI: 10.11996/JG.j.2095-302X.2021050744

• 图像处理与计算机视觉 • 上一篇    下一篇

适应性增强胶囊网络的古壁画朝代识别算法

  

  1. 1. 太原科技大学计算机科学与技术学院，山西 太原 030024；2. 忻州师范学院计算机系，山西 忻州 034000
  • 出版日期:2021-10-31 发布日期:2021-11-03

A dynasty classification algorithm of ancient murals based on adaptively enhanced capsule network 

  1. 1. School of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan Shanxi 030024, China; 2. Department of Computer Science and Technology, Xinzhou Teachers University, Xinzhou Shanxi 034000, China
  • Online:2021-10-31 Published:2021-11-03

PDF (PC)

93

摘要： 针对壁画图像自身存在多义性、不同朝代的壁画绘画风格存在差异性和传统的人工直接参考壁 画文本或历史文献识别壁画朝代存在费时、费力等问题，提出了适应性增强胶囊网络(AEC)的古壁画朝代识别 算法，对莫高窟壁画自动进行朝代识别。在原生胶囊网络基础上，首先引入预先卷积结构对壁画图像进行高层 特征提取；其次增加均层激活增强模型的拟合性能；最后对胶囊网络进行适应性增强，在提高梯度平滑度的基 础上利用自适应学习率进行优化提高模型的分类精度。实验结果表明在所构造的 DH1926 壁画数据集上，AEC 模型准确率为 84.44%、平均精确度(MAP)为 82.36%、平均召回率(MAR)为 83.75%、综合评价指标为 83.96%。 与改进的卷积神经网络(CNN)和原生胶囊网络等其他网络结构相比，各项评价指标均有 3%以上的提升，有较 强的拟合性能，能够多层次提取壁画的丰富特征，表达图像更细节的语义信息，在莫高窟壁画朝代识别中具有 更高的准确度和更好的鲁棒性，有一定的应用价值与研究意义。

关键词: 预先卷积, 均层激活, 适应性增强, 胶囊网络, 壁画朝代分类

Abstract: In view of the ambiguity of mural images, the differences of mural painting styles in different dynasties, and the high time-consumption and difficulty of the direct traditional manual reference to mural texts or historical documents to identify mural dynasties, an adaptive enhancement capsule network (AEC) algorithm for ancient mural dynasties identification was proposed to automatically identify the dynasties of Mogao Grottoes murals. Based on the original capsule network, the pre-convolution structure was introduced to extract the high-level features of mural images. Secondly, the fitting performance was increased for homogeneous layer activation enhancement model. Finally, the adaptability of the capsule network was enhanced. On the basis of the improved gradient smoothness, the adaptive learning rate was employed to optimize the model, thus improving the classification accuracy of the model. The experimental results show that on the constructed DH1926 mural data set, the accuracy rate of the adaptively enhanced capsule network model is 84.44%, the average accuracy (MAP) is 82.36%, the average recall rate (MAR) is 83.75%, and the comprehensive evaluation index is 83.96%. Compared with other network structures, such as improved convolutional neural network (CNN) and native capsule network, each evaluation index has been improved by more than 3%, and displayed strong fitting performance. It can extract rich features of murals at multiple levels and express more detailed semantic information of images. It is advantageous in higher accuracy and better robustness in the dynasty recognition of Mogao Grottoes murals, and is of certain application value and research significance. 

Key words:  , pre-convolution, homogeneous layer activation, adaptive enhancement, capsule network, mural dynasty classification 

中图分类号: