Research And Implementation Of Image Recognition Model Online Fast Training S Ystem For Small Scale Data

In recent years,whether it is the shift of STEM’s hot spots in primary and secondary schools to AI at home and abroad,or the introduction of content related to AI in online education platforms,shows that global AI education is showing a trend of younger age.At present,the machine learning platform for teenagers at home and abroad is still in the primary stage of development,and there are many problems such as low training efficiency and insufficient data scale.Among all kinds of training data,the quality of image data and algorithm selection have the most obvious influence on the model training effect.As an important part of machine learning platform,image classification model has the characteristics of high data quality,high training cost and long training cycle,which can’t meet the needs of training the model quickly with insufficient data.Therefore,this thesis designs an image recognition model online fast training system for small scale data,and makes an in-depth study on the automatic generation technology of image models.Firstly,this thesis analyzes and summarizes the research results of automatic machine learning and transfer learning in the field of small-scale data image classification at home and abroad.Secondly,aiming at the problem of data scale and distribution,this thesis designs a dynamic data enhancement algorithm based on data balance,and proposes an enhancement strategy reuse optimization technology based on tag information and Bayesian algorithm to realize dynamic data enhancement.In order to realizing the automatic training function,an adaptive image classification model is proposed.Firstly,an adaptive algorithm framework based on transfer learning is designed for the model.Secondly,this thesis designs an automatic model selection algorithm based on task complexity.Finally,the dynamic data enhancement process is combined with the model automatic training process,and the automatic model generation solution AutoTrain is designed to realize the synchronous optimization of model parameters and enhancement strategy.Based on the above theoretical research results,this thesis designs,implements and tests each module of the online fast model training system for small-scale data image classification of machine learning education platform,and compares it with the current mainstream automatic machine learning system.Experimental results show that the dynamic data enhancement technology and the adaptive image classification model can effectively improve the training efficiency of user model,reduce the requirements of computing resources,and solve the problem of the models’bad performance caused by uneven data distribution.