基于图像模态迁移与集成的中国陶瓷派系识别模型构建研究

Identifying Chinese Ceramic Genres Based on Image Modal Transfer and Ensemble Learning

[目的]构建中国陶瓷图像派系识别模型,实现对陶瓷图像派系的自动分类和识别,为陶瓷文化研究和数字化保护提供技术支撑。[方法]采用“端到端范式”构建模型,将迁移学习、集成学习应用到陶瓷派系识别中,并利用DCGAN算法进行样本平衡,根据各种陶瓷品的工艺和艺术风格,实现对10个陶瓷派系的识别与分类。[结果]基于端到端范式构建的陶瓷派系识别模型能够很好地提取图像特征并完成识别任务,且效果优于手工设计特征工程的基线模型。迁移学习使得预训练模型学习到的特征可以有效迁移到陶瓷派系识别这一细粒度的下游任务中,最优模型准确率达到73.16%;改进的Stacking集成方法融合上述模型学到的知识,最终准确率达到81.39%。[局限]本文所使用的数据来源于百度图片,数据来源较为单一,对模型的性能产生一定影响。[结论]基于迁移学习与集成学习的端到端图像模态分类模型能够有效地应用到陶瓷这一细粒度任务中,取得了较好的效果。

[Objective]This paper constructs a clique recognition model for Chinese ceramic images.It aims to automatically classify and recognize the clique of ceramic images and provide technical support for the research and digital protection of ceramic culture.[Methods]We adopted the“end-to-end learning”paradigm to build the new model.It applied transfer learning and ensemble learning technology to ceramic cliff identification.We also used the DCGAN algorithm to balance samples.We examined the new model with ten cliques of ceramics based on their types,crafts,and artistic styles.[Results]The proposed model could more effectively extract ceramic image features and recognize ceramic cliques than the baseline models with manually designed feature engineering.Transfer learning enables the extracted features to be effectively transferred to the fine-grained downstream tasks.The accuracy of the new model reached 73.16%.The improved Stacking method integrated knowledge from the proposed models and increased the final accuracy to 81.39%.[Limitations]The data used in this paper is from Baidu pictures,which need to be expanded to improve the model’s performance.[Conclusions]The new model could effectively classify and identify ceramic images.