Research On Deep Learning Recognition Algorithm And Application Based On Small Sample Data Enhancement Technology

Deep learning technology is one of the hot spots of machine learning research and application in the field of scientific computing and image processing at present.The effective use of machine learning includes three aspects: data,algorithm and computing power.Now,research and application mainly focus on algorithms.However,the main constraint on the effect of machine learning application is the data for many important applications.Especially for some application fields with less data volume and insufficient data features,the quality and scale of data cannot match the requirements of algorithms.To solve the problem of poor training effect or failure of deep learning model due to poor data quality and small data scale,data enhancement methods include data transformation enhancement,model generation and so on are currently commonly used.In this thesis,numerical simulations and physical experiment simulations are proposed to enhance small sample data.We analyze data features and the generation conditions of small sample data,and then design theoretical and experimental methods to retain practical features and enhance data quality and scale for producing high-throughput data from small samples.Generate a new data set as the similar data of the actual data,namely the simulation data.Taking the quantum piezoelectric electronic devices in the frontier intersection of quantum physics and future electronics as an example,aiming at the numerical simulation data in scientific research,the numerical calculation data of quantum devices are used as the simulation data in combination with the data scale and the regularity of data characteristics.We have explored the application of deep learning technology in quantum device data calculation.Topological materials are the current research frontiers of quantum piezoelectric electronics,the change of topological insulating state of materials can be induced in quantum piezoelectric materials through the piezoelectric electric field caused by spontaneous polarization with almost no lattice mismatches.In this thesis,the studied system is the topological insulator composed of Hg Te/Cd Te quantum piezoelectric materials.The deep learning network is trained by using the data of quantum piezoelectric electronics,such as electron density distributions and conductance which are calculated by the numerical simulation software KWANT.The unknown data can be predicted through the trained network and R-squared is 0.999 of the conductance from KWANT and deep learning.Aiming at the data generated in real physical scenes such as the research of the small sample digital text character recognition project with small data scale and poor data quality.We have analyzed the physical scene of the data generation and simulated the characteristics of the small sample data.Finally,we have designed the high-throughput data generation experiment to generate the simulated data.The use of simulation data training network can strengthen the main features and filter out noise in the deep learning network training which can realize the rapid deployment of applications in important fields.The recognition accuracy for the real data is more than 99% in practical engineering application.