From classification to matching: A CNN-based approach for retrieving painted pottery images

Recently artificial intelligence has begun to assist archaeologists in processing images of archaeological artifacts. We report a convolutional neural network approach to obtain feature vectors of painted pottery images by a preliminary classification machine learning of the cultural types. The model, trained on a photographic image dataset of Chinese Neolithic color-painted pottery, achieved 92.58% precision in assigning vessel images to corresponding archaeological types. The feature vectors contain information of vessel shape, color, and ornamentation design, based on which similarity coefficients for the images in the dataset were calculated. The quantitative measurement of similarity allows searching for the closest match to artefacts in the dataset, as well as a network of vessels in terms of similarity. This work highlights the potential of CNN approaches in curating of archaeological artifacts, providing a new tool assisting to study chronology, typology, decoration design, etc.

近年来，人工智能已开始协助考古学家处理考古文物图像。本研究提出一种卷积神经网络（Convolutional Neural Network, CNN）方法，通过对文化类型开展初步机器学习分类，提取彩陶图像的特征向量。本模型基于中国新石器时代彩陶摄影图像数据集训练完成，在将器物图像归类至对应考古类型的任务中达到了92.58%的精确率。该特征向量涵盖器物器形、色彩与装饰纹样的相关信息，基于此可计算数据集内各图像的相似度系数。这种量化的相似度度量方式，既支持在数据集内检索与目标文物最匹配的样本，也可构建基于相似度的器物关联网络。本研究凸显了卷积神经网络方法在考古文物整理工作中的应用潜力，为考古年代学、类型学、装饰纹样研究等领域提供了全新的辅助研究工具。