Microfossils Image Few Shot Recongition Based On Deep Residual Network And Transfer Learning

The southern part of Shaanxi Province,China is rich in microfossils from the Kuanchuanpu Formation of the Early and Middle Cambrian(535 million years ago),rich in tetragonal tower shells,early animal embryos,protoconodonts and other microfossils.It is a very important research direction to help study the origin of animals in the Cambrian period and the cause of the explosion.However,early researchers mainly relied on manual microscope screening to find micro-fossils due to technical limitations.The number of microfossils is huge,but because of the scarce samples of research value,the error of artificial methods,and the low efficiency of discovery,it greatly affected the progress of early life research in the Cambrian.Based on the above research background,this paper establishes a micro-fossil image data set of the Kuanchuanpu Formation in southern Shaanxi,and proposes two solutions,machine learning and deep residual network,for the classification and recognition of small samples of micro-fossils.The research is as follows:(1)In response to the lack of public microfossil data sets,the Kuanchuanpu microfossil data set in southern Shaanxi was established.Based on the collected fossil samples,through the process of acid bubble,shooting and classification,the establishment of the Kuanchuan ball,A total of4507 micro-fossil image datasets in 9 categories including Xixiang Tower and Conodont.(2)Based on machine learning direction gradient histogram,Gabor transform and other methods to identify small samples of micro-fossils,two methods of microscopic image data expansion and CT image multi-angle projection are proposed.The average recognition rate of the nine types of microfossil microscopic expanded images reached 93%.In the case of very few samples,the recognition rate of micro-fossil images using multi-angle projection of CT images has increased by more than 40% compared with micro-fossil micro-image data.(3)Micro-fossil small sample image recognition method based on the combination of migration learning and Res Net residual network.This method uses a 34-layer deep residual network as a training model,adopts a model-based migration method,freezes low-level parameters,and is retraining On the upper network parameters,a micro-fossil image recognition model is constructed,and the Tr Ada Boosting algorithm is used as a method to optimize the parameters of the micro-fossil model.A total of 15 different experiments were designed.Experiments show that the accuracy of the deep residual network fossil recognition model proposed in this paper has reached 96.7%,which is better than support vector machines and random forests.And for a small sample(1-shot)image recognition rate of 70.5%,it is better than traditional machine learning methods,VGG,and Goog Le Net.The method proposed in this paper has a recognition rate of more than 95% for the microfossils in Kuanchuanpu,southern Shaanxi,and reached 70.5% in the case of a single sample.It has reached the research goal of this paper and has certain research significance.