为了探索基于深度神经网络模型的牙形刺图像智能识别效果, 研究选取奥陶纪8种牙形刺作为研究对象, 通过体视显微镜采集牙形刺图像1188幅, 收集整理公开发表文献的牙形刺图像778幅, 将图像数据集划分为训练集和测试集。通过对训练集图像进行旋转、翻转、滤波增强处理, 解决了训练样本不足的问题。基于ResNet-18、ResNet-34、ResNet-50、ResNet-101、ResNet-152五种残差神经网络模型, 采用迁移学习方法, 对网络模型进行训练以获取模型参数, 五种模型测试Top-1准确率分别为85.37%、85.85%、83.90%、81.95%、80.00%, Top-2准确率分别为94.63%、94.63%、94.15%、93.17%、93.66%, 模型对牙形刺图像具有较好的识别效果。通过对比研究发现, ResNet-34识别准确率最高, 说明对于特征简单的牙形刺属种, 增加网络深度并不一定能提升准确率, 而确定合适深度的模型则不仅可以提高识别准确率, 还可以节约计算资源。通过ResNet-34模型的迁移学习训练和重新训练效果对比可以看出, 迁移学习不仅可以获得较高的准确率, 而且可以较快获取模型参数, 因而可作为小样本古生物化石图像识别的重要方法。研究还发现, 体视显微镜下牙形刺图像的识别准确率高于扫描电镜下图像识别准确率, 化石完整性和相似性、照相角度以及数据集的大小是影响图像识别准确率的主要原因。

This paper aims to explore the image recognition effect of conodonts based on deep neural network model. Eight kinds of Ordovician conodonts were selected as the research objects. 1188 photos of conodonts were taken under stereomicroscope, and 778 SEM images of conodonts were collected from published papers. The image data set was divided into training set and test set. In order to solve the problem of insufficient training samples, the methods of data augmentation of rotating, flip-ping, and filtering are used. Five residual neural network models of resnet-18, resnet-34, resnet-50, resnet-101 and res-net-152, were used to train to obtain the transfer learning model parameters. The top-1 test accuracy for the five models is 85.37%, 85.85%, 83.90%, 81.95%, and 80.00% respectively. The top-2 test accuracy is 94.63%, 94.63%, 94.15%, 93.17%, and 93.66% respectively. It was found that resnet-34 recognition has the highest test accuracy, which indicates that for the simple fossil like conodonts, increasing the depth of network does not improve the test accuracy. Selecting the appropriate depth model can not only improve the recognition accuracy, but also save computing resources. By comparing the transfer learning and retraining results of resnet-34 model, it was found that transfer learning can not only obtain high accuracy, but also quickly obtain model parameters, which is an important method for small sample fossil image recognition. The recogni-tion accuracy of conodont image under stereomicroscope is higher than that under scanning electron microscope. Integrity and similarity of conodont, camera angle, and dataset size are the main factors that affects the image recognition accuracy.