化石的系统分类是古生物学研究的基础工作。传统的古生物系统分类对专业经验要求高, 多重复劳动费时费力, 并且不同分类专家之间存在鉴定差异。本文结合迁移学习与卷积神经网络技术, 将其引入基于图像的古生物系统分类工作中, 探索其辅助古生物学家进行高效、准确分类的有效性。本文构建了一个包含5个纲及亚纲的无脊椎动物化石类群共3358张化石图像的数据集, 整体具有样本总量有限、形态特征维度复杂、组间样本分布不均衡等特点。针对这些特点, 本研究选取了VGG16模型结合迁移学习的方法进行训练, 得到的最优模型在测试集上的分类准确率达到97.33%, 精确度达到97.26%, 召回率达到97.07%, F1分数达到97.16%。这一结果表明, 本文建立的基于迁移学习的卷积神经网络化石图像分类流程, 包括在数据预处理中使用数据增强, 在训练中使用具有针对性的微调, 均可以有效地改善使用小数据集训练复杂深度模型时的局限性。面对复杂度较高、数据量较小的科学数据集, 选择合适的迁移学习训练策略, 也可以在有限的计算成本内, 训练出满足预期性能的高性能模型。

Taxonomic identification of fossil is fundamental in paleontology research. However, traditional taxonomic methods rely heavily on expert experience and are inherently time-consuming. The development of deep learning architectures has made image-based fossil identification increasingly feasible. In this study, we apply an integrated framework that combines convolutional neural networks (CNN) with transfer learning to develop well-designed and trained machine-learning models for paleontological research. A fossil image dataset comprising 3, 358 images in five invertebrate classes and subclasses is compiled to train a VGG16 classifier. This dataset presents challenges due to its limited dataset size, high-dimensional data features, and inter-class/subclass sample imbalances. To address these issues, data augmentation and a two-phase transfer learning strategy are employed in the training process. In the first phase, only the final fully connected layer is randomly initialized and trained, while other layers are frozen with pretrained ImageNet weights. In the second phase, the whole model is fine-tuned with a reduced learning rate. The optimized model achieves an accuracy of 97.33%, a precision of 97.26%, a recall of 97.07%, and an F1-score of 97.16% on the test set. An ablation study is also conducted to test the effectiveness of transfer learning. A VGG16 classifier trained with random initialization achieves an accuracy of 74.89% on the test set. These results demonstrate that the constrained scientific datasets can be effectively utilized through (1) transfer-learning-enhanced CNN architectures; (2) strategic data augmentation protocols and (3) a progressive fine-tuning process. Notably, the model successfully accommodates heterogeneous input conditions of the fossil images, including varying imaging equipment, taxonomic diversity, complex image backgrounds, and environmental noises. This highlights the potential of the proposed framework for developing high-precision classifiers capable of fine-grained taxonomic discrimination at genus or species level, while maintaining computational efficiency. The achieved performance benchmarks suggest promising applications for automated fossil identification systems in paleontological research.

江苏省重大科技开放合作平台建设项目(BZ2023068)、创新能力建设计划项目(BM2021807)和中央高校基本科研业务费专项资金(关键地球物质循环前沿科学中心“GeoX”交叉项目0206-14380121)联合资助