一种改进的SAM-ShuffleNet明青花龙纹断代模型  

An improved SAM⁃ShuffleNet model for dating Ming blue⁃and⁃white dragon patterns

作　　者：赵谦[1] 屈佳伟 ZHAO Qian;QU Jiawei(College of Communication and Information Technology,Xi’an University of Science and Technology,Xi’an 710054,China)

机构地区：[1]西安科技大学通信与信息学院,陕西西安710054

出　　处：《现代电子技术》2025年第9期116-123,共8页Modern Electronics Technique

基　　金：西安市科技计划(24KGDW0030);陕西省科技成果转化计划-百项科技成果转化行动项目(2023-YD-CGZH-29);陕西秦创原“科学家+工程师”队伍建设项目(S2023-YF-LLRH-QCYK-0271)。

摘　　要：青花瓷作为瓷器中的明珠,其独特的纹饰承载着深厚的文化底蕴。在众多青花瓷纹饰中,明代龙纹颇具代表性,对明代前、中、后三期的青花瓷龙纹进行断代识别有着极高的历史文化研究价值。文中针对明代龙纹断代问题提出一种改进的深度学习模型,该模型引入空间注意力机制(SAM)改进原网络ShuffleNetV2,能够捕获图像中的长距离依赖关系,用于增强模型的表达能力,并利用迁移学习策略、交叉熵损失函数和RMSProp自适应学习率优化器在青花瓷龙纹数据集上进行实验验证,用于准确年代鉴定。实验表明,在迭代100次、批次大小为16、学习率为0.001以及权重衰减为10-5的条件下,明代前、中、后期断代总体准确率为95.67%,召回率为95%,F1-Score为95%。与VGG16、ResNet18等8个模型对比,所提模型在识别准确率上显著提高,运算速度约是ResNet18的4.9倍,是VGG16的5.9倍,这证明该模型不仅实现了明代青花瓷龙纹的精确断代,还具备快速识别、体积紧凑的特点,为文化遗产保护和鉴定提供了新的技术方法。Blue-and-white porcelain,as the pearl of ceramics,carries profound cultural heritage with its unique patterns.Among the numerous patterns of the blue-and-white porcelain,the dragon patterns in the Ming Dynasty are quite representative,and the identification of the dragon patterns in the early,middle,and late Ming Dynasty has extremely high historical and cultural research value.In view of the above,an improved deep learning model is proposed for the dating of the dragon patterns in the Ming Dynasty.In the model,a spatial attention mechanism(SAM)is introduced to improve the original network ShuffleNetV2,which can capture long-distance dependencies in images and enhance the presentation ability of the model.By using transfer learning strategies,cross-entropy loss function,and the RMSProp adaptive learning rate optimizer,the model was experimentally verified on a dataset of blue-and-white porcelain dragon patterns for accurate dating.The experiments show that under the conditions of 100 iterations,a batch size of 16,a learning rate of 0.001,and a weight decay of 10-5,the overall accuracy rate for dating the dragon patterns in the early,middle,and late Ming Dynasty is 95.67%,with a recall rate of 95%and an F1-Score of 95%.In comparison with the eight models including VGG16 and ResNet18,the proposed model improves recognition accuracy significantly,operates 4.9 times faster than ResNet18,and 5.9 times faster than VGG16.This proves that the model not only achieves precise dating of the blue-and-white porcelain dragon patterns in the Ming Dynasty,but also possesses the characteristics of fast recognition and compact size,providing new technological methods for the protection and identification of cultural heritage.

关 键 词：龙纹断代 深度学习 ShuffleNet 空间注意力机制 RMSProp 迁移学习策略 

分 类 号：TN911.73-34[电子电信—通信与信息系统]