Research On Automatic Classification And Image Semantic Segmentation Of Fossil Based On Deep Learning

Paleontology provides the most direct evidence for the origin and evolution of life.Fossils that have witnessed the history of life on Earth are the main research objects of paleontology.They are characterized by large numbers and rich types.Therefore,the fossil data analysis is time-consuming and requires strong professional skills.With the rapid development of deep learning(DL)technology,paleontological experts have also begun to apply DL technology to classify and semantically segment fossil images.However,2D images can only provide limited information on fossils.They cannot reveal the special morphological structure,which may lead to some misidentification of fossils.Threedimensional(3D)imaging technology,especially micro-computerized tomography(CT)technology,is widely applied in the digitalization of fossils.Simultaneously,a large number of digital 3D data of precious fossils are rapidly accumulated.Paleontologists urgently need shared storage space and professional analysis platform for fossil 3D data.Therefore,our research involves the following three aspects about 3D data management,CT image semantic segmentation,and 3D model automatic classification of microfossils:1.This paper designed and developed a data platform “Archive of Digital Morphology(ADMorph)” with fossil morphological data as the primary metadata supported by the “Earth Big Data Science Project”.It coverd almost all categories in vertebrate paleontology,including fossil fish,amphibians,reptiles,birds,mammals,and humans.This paper provided an open source dataset of fossil 3D models(ADMorph)using micro CT and image processing technology.During the process,this research was mainly responsible for CT data acquisition of fossils and the creation of most 3D models.Currently,the dataset has contained over10,000 3D models of bone fragments,teeth,and scales from five major paleo-fish groups.2.This paper proposed a DL-based image semantic segmentation method for microfossil CT data.This paper applied an improved UNet network with Res Net34 model as the backbone and fused multi-level feature with feature pyramid at different scales.It could extract features to perform pixel-level image semantic segmentation.The method achieved the highest global Io U score of 98.03% and DSC score of 98.84% on 6 sets of experimental dataset.Comparing the proposed method with the traditional image binarization method and UNet network,the global Io U score and DSC score was respectively higher than(10.14% and5.40%)and(7.11% and 5.72%).3.This paper proposed a DL-based classification method for microfossil 3D models.This method included data preprocessing(DP),feature extraction,feature fusion,and 3D model classification.It achieved a macroaveraging accuracy of 97.60%,the highest 100%(Psarolepis,Guiyu),and the lowest 88.78%(Parathelodus liaokuoensis)on the experimental dataset.Comparing the proposed method with the classic classification methods of 3D models(Vox Net,Point Net,and MVCNN),the macroaveraging accuracy was higher than19.45%,33.47%,and 6.06%,respectively.At the same time,the ablation experiment was carried out.Using DP model,the macroaveraging accuracy has been increased by 4.23%.Using deep neural network(DNN)and support vector machine(SVM)model,the macroaveraging accuracy has been increased by 5.56%.In summary,this paper focuses on the creation of 3D model dataset and the research on key technologies of semantic segmentation and automatic classification,taking paleo-fish microfossil data as the main research object.DL-based fossil data analysis method proposes a new idea for the research on fossil morphology.Compared with the traditional methods,this method has greater efficiency and objectivity.It will give hope for the determination of geological age,the division and comparison of strata,and the search for mineral resources.