| Various phosphatized microfossils from the early Cambrian of southern Shaanxi are of high research value,and they are important windows for exploring the Cambrian explosion of life and the origin of animal phyla.However,microfossil samples,which provide key clues for understanding the evolutionary history of animals,are extremely scarce,or even single,and are easily overwhelmed by the dominant species with higher abundance,resulting in a very inefficient manual selection of such fossils under the microscope,and the existing artificially intelligent fossil recognition models are exceedingly dependent on a large number of training samples,and the recognition accuracy will plummet once the sample size is reduced.In response to the above situation,this master’s thesis presents an in-depth study of the microfossil recognition problem under one-shot conditions by constructing a microfossil image dataset of the Cambrian Kuanchuanpu Formation,with the aim of discovering and analyzing the biological properties of more rare microfossils to promote the research progress of micropaleontology.The main work is described as follows:(1)Research on One-Shot Microfossil Recognition Method Based on Optimized Siamese Network Structure(OSIAM).OSIAM consists of two parts: first,a lightweight siamese network for one-shot microfossil recognition is designed by optimizing the sub-network structure of the siamese network based on the characteristics of microfossil images;second,due to the lack of sample size of microfossils,a stratified stochastic gradient descent algorithm that allows updating the network parameters layer by layer during the training phase is proposed in addition to expanding the training data size by using random pairing and data augmentation.Experimental results show that the OSIAM method has the lowest number of parameters and moderate training time compared to using Let Net5 and VGG16 as sub-networks,without relying on high-performance computing equipment;meanwhile,the OSIAM method can quickly converge to the optimal model and effectively recognize one-shot microfossils with an average accuracy of 70.7%.(2)Research on One-Shot Microfossil Recognition Method Based on Siamese Network and Transfer Learning(TLSIAM).To address the problem that the number of categories has a greater impact on the performance of OSIAM,we introduce the idea of transfer learning on the basis of OSIAM.A pre-trained model is first constructed using the TLL(Totally-LooksLike)dataset,and then the parameters of the pre-trained model are fine-tuned by selecting a suitable transfer strategy and using the automatic weight adjustment mechanism of Tradaboost to finally obtain a highly accurate and robust one-shot microfossil recognition model.Experimental results show that the average classification accuracy of the TLSIAM method for recognizing different classes of one-shot microfossils reaches 83.9%,which is13.2% higher than that of the OSIAM method,and outperforms various baseline methods. |
