Research On Image Recognition Based On Machine Learning

Convolution neural network is a branch of machine learning research field,which is a new field developed based on artificial neural network.With the development of big data depth model and hardware equipment,deep learning technology has also been widely developed to promote the development of computer vision,speech recognition,natural language processing and other fields,and the success of these models is depend on a large amount of data,however.In practice,due to the high cost and time-consuming collection of data sets and training models,and the limited data set,resulting in the over-fitting of the trained models,this thesisbased on the open kaggle dataset in depth residual neural network.The accuracy of validation dataset is 91.04%,and the accuracy of dataset in depth residual neural network based on engineering practice is 85.59%.The precision of training and verification is not enough in engineering practice,resulting in serious over-fitting phenomenon.Therefore,this paper presents two methods to reduce over-fitting based on the residual neural network.The first method proposes a method of reducing the number of convolution kernels by reducing the number of convolution kernels by reducing the number of convolution kernels,Over-fitting and cross-combination method make the accuracy of kaggle data of dogs and cats on validation dataset reach 95.78% and 87.64% on dataset of 31 types of engineering practice.The second method is based on Finetune residual neural network On the basis of Finetune residual neural network method,the precision of model was improved.The precision of Fintune residual network was 99.42% on kaggle cats and dogs verification datasets,and the accuracy of datasets was verified in 31 kinds of engineering practice.The accuracy of datasets was 99.19% The Recurrent-Finetune residual neural network approach achieves 99.58% accuracy in kaggle cats and dogs validation datasets and 99.48% in 31 validation datasets for engineering practice.