基于深度学习的介形类化石层次化识别

Hierarchical recognition of ostracod fossils based on deep learning

介形类化石对地质年代的确定、古湖泊和古海洋的研究、古环境的重建以及海底石油资源的勘探等工作都具有重要意义。然而,现有识别化石颗粒的方法费时费力,准确率也有待提高。鉴于介形类化石颗粒的类别具有科、属、种的层次结构,种类数量庞大,所以笔者等提出了一种层次化识别方法。首先进行目标检测,实现介形类化石的定位与属类划分;之后在目标检测模块的基础上进行智能识别,使用卷积神经网络和支持向量机提取属类下更细微的种类特征,实现化石种类划分。实验结果表明,笔者等提出的分层次识别模型能检测出化石图像中所有化石颗粒的位置信息并对其进行分类,分类准确率可达95%,且相较于未进行分层次识别的模型,能将识别准确率提升1.8%~5.8%。

Objectives:The research of ostracod fossils is of importance to the determination of geological age, the study of paleo-oceans, the reconstruction of paleo-environments and the exploration of submarine oil resources. However, the existing methods for identifying fossil are time-consuming and labor-intensive, and the accuracy rate needs to be improved. In this paper, we propose a hierarchical recognition method due to the hierarchical structure of ostracod fossils’ categories(families, genera, and species) and wide range of species.Methods: First, perform object detection to realize the positioning and genus-level classification;then intelligently identification based on the object detection module, uses CNN to extract features among same genus particle and SVM for species-level classification.Results:It is demonstrated that the hierarchical identification can locate and classify particles in fossil images and our test presents the 95% classification accuracy. The accuracy increases by 1. 8% ~5. 8% which are compared with non-hierarchical recognition one.Conclusions: In this paper, the accuracy of the proposed method reaches 95%, which confirms the feasibility and application prospects of computer vision methods based on deep learning in paleontology research. The computer directly gets the paleontological image feature through learning and classify automatically, making full use of the computer’s active learning characteristics. In the future research work, we will further expand the ostracod fossil database, the data samples and categories to improve the model of recognition accuracy, generalization and applicability. Based on the proposed model, an intelligent ostracod fossils identification system will be developed to improve the efficiency of ostracod fossil identification.