Xu Jiahui, Lu Yang, Xu Xu
Jilin Normal University Siping City Jilin Province China.
Ecol Evol. 2025 Aug 3;15(8):e71845. doi: 10.1002/ece3.71845. eCollection 2025 Aug.
Traditional fossil identification relies on the rich experience and knowledge of paleontologists, and existing intelligent identification methods mainly rely on deep learning to train on a large number of fossil graphic samples to achieve a high degree of precision. In order to solve the above problems and still be able to accurately recognize small samples of rare fossils, we try to use the generative adversarial network (GAN) combined with ResNet50, EfficientNet, and customized CNN architectures, which are applied to the identification of small samples of fossils. First of all, the generator of GAN is fully trained, using it to generate a large number of samples to expand the dataset, enriching the image features extracted by the model, and then through the neural network to analyze the image abstraction computation, and finally, the best fossil identification model is trained through multiple iterations. Using the method of this paper on the same dataset with a data enhancement method for comparison experiments, the experimental results show that the accuracy rate reaches 93% in the case of epochs 100, higher than the other experimental results, and has a significant advantage in the recognition of fossils with scarce samples.
传统的化石鉴定依赖于古生物学家丰富的经验和知识，现有的智能鉴定方法主要依靠深度学习，在大量化石图形样本上进行训练以实现高精度。为了解决上述问题，同时仍能够准确识别珍稀化石的小样本，我们尝试使用生成对抗网络（GAN）结合ResNet50、EfficientNet和定制的卷积神经网络（CNN）架构，将其应用于化石小样本的鉴定。首先，对GAN的生成器进行充分训练，用其生成大量样本以扩充数据集，丰富模型提取的图像特征，然后通过神经网络进行图像抽象计算分析，最后通过多次迭代训练出最佳的化石鉴定模型。将本文方法与一种数据增强方法在同一数据集上进行对比实验，实验结果表明，在100个轮次的情况下准确率达到93%，高于其他实验结果，并且在样本稀缺的化石识别方面具有显著优势。
