基于ResNet18的坭兴陶识别研究
Research on Nixing Pottery Recognition Based on ResNet18
摘要: 坭兴陶是中国的四大名陶之一，其外观与景德镇瓷、德化瓷、宜兴紫砂高度趋同，肉眼难以分辨，不了解的人往往容易将其混淆。由此种种，制约了其文化的传播与相关产业的发展。本文针对坭兴陶与相似陶器外观易混淆、鉴别难的核心问题，提出一种基于ResNet18的坭兴陶细粒度识别方法。采用困难样本反馈策略(困难样本过采样 + 分层特征微调)对ResNet18模型进行改进，并(随机)结合MixUp和CutMix算法的数据增强技术(线性加权混合样本和标签以及通过裁剪拼接不同样本区域等)提升模型的泛化能力。实验结果表明，传统的ResNet18模型在测试集上的整体准确率为82.22%，而改进后的ResNet18模型在测试集上的整体准确率可达84.44%，与EfficientNet-B0、MobileNetV2、VGG16等主流模型相比，其在准确率与模型轻量化之间实现更优平衡。研究为坭兴陶的智能鉴别与非遗数字化保护提供了可行的方向和高效的技术支撑。
Abstract: Nixing Pottery is one of China’s four great famous pottery. Its appearance is highly similar to Jingdezhen Porcelain, Dehua Porcelain, and Yixing Zisha Pottery, making it difficult to distinguish with the naked eye. People who are not familiar with them often confuse them. All these issues have restricted the dissemination of its culture and the development of related industries. Focusing on the core problem that Nixing Pottery is easily confused with similar pottery in appearance and difficult to identify, this paper proposes a fine-grained recognition method for Nixing Pottery based on ResNet18. The ResNet18 model is improved by adopting a hard sample feedback strategy (hard sample oversampling + hierarchical feature fine-tuning), and data augmentation technologies of MixUp and CutMix algorithms (linearly weighted mixing of samples and labels, as well as cropping and splicing regions of different samples, etc.) are (randomly) combined to enhance the generalization ability of the model. Experimental results show that the overall accuracy of the traditional ResNet18 model on the test set is 82.22%, while the overall accuracy of the improved ResNet18 model on the test set can reach 84.44%. Compared with mainstream models such as EfficientNet-B0, MobileNetV2, and VGG16, it achieves a better balance between accuracy and model lightweighting. This research provides a feasible direction and efficient technical support for the intelligent identification of Nixing Pottery and the digital protection of intangible cultural heritage.
文章引用：苏庆鸥, 黄媛, 钟畅, 刘柏霆. 基于ResNet18的坭兴陶识别研究[J]. 计算机科学与应用, 2025, 15(10): 266-275. https://doi.org/10.12677/csa.2025.1510266

参考文献

[1] 汤博宇, 焦良葆, 徐逸, 等. 基于改进ResNet-50的图像特征提取网络[J]. 计算机测量与控制, 2023, 31(6): 162-167.
[2] Chowdhury, M., Shrima, S.R. and Islam, M.S. (2025) Comparative Analysis of Deep Learning Architectures for Multi-Class Mineral Classification: A Study Using EfficientNet and ResNet Models. Earth Science Informatics, 18, Article No. 485. [Google Scholar] [CrossRef] 
[3] 罗鑫, 王艳艳, 刘学渊, 等. 基于深度学习的林地澳洲坚果检测识别[J]. 森林工程, 2023, 39(2): 113-120.
[4] Yang, L., Zhou, W. and Qiu, W. (2025) Chronological Classification of Ming and Qing Dynasty Ceramics Images Based on an Enhanced ResNet50 Model. STAR: Science & Technology of Archaeological Research, 11, e2498260. [Google Scholar] [CrossRef] 
[5] Jurj, S.L., Opritoiu, F. and Vladutiu, M. (2018) Identification of Traditional Motifs Using Convolutional Neural Networks. 2018 IEEE 24th International Symposium for Design and Technology in Electronic Packaging (SIITME), Iasi, 25-28 October 2018, 191-196. [Google Scholar] [CrossRef] 
[6] Gao, L., Zhang, X., Yang, T., Wang, B. and Li, J. (2023) The Application of ResNet-34 Model Integrating Transfer Learning in the Recognition and Classification of Overseas Chinese Frescoes. Electronics, 12, Article 3677. [Google Scholar] [CrossRef] 
[7] He, K., Zhang, X., Ren, S. and Sun, J. (2016) Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, 27-30 June 2016, 770-778. [Google Scholar] [CrossRef] 
[8] 邢俊杰, 贾民平, 许飞云, 等. 基于CutMix和YOLOv3的工件表面小缺陷识别方法(英文) [J]. 东南大学学报(英文版), 2021, 37(2): 128-136.
[9] 陆健强, 林佳翰, 黄仲强, 等. 基于Mixup算法和卷积神经网络的柑橘黄龙病果实识别研究[J]. 华南农业大学学报, 2021, 42(3): 94-101.
[10] Liu, Y., Yang, X., Zhou, S., Liu, X., Wang, Z., Liang, K., et al. (2023) Hard Sample Aware Network for Contrastive Deep Graph Clustering. Proceedings of the AAAI Conference on Artificial Intelligence, 37, 8914-8922. [Google Scholar] [CrossRef]