Study On Migration Model Based On Convolutional Neural Network

Although the convolutional neural network exhibits excellent performance in target recognition,segmentation and detection,the huge demand for the number of samples in network training severely limits its application in certain specific situations.In order to better solve this problem and expand the convolutional neural network,migration learning has been paid more and more attention by researchers to convolutional neural networks.Based on this,this thesis carried out research work on the migration model based on convolutional neural network,mainly related to the following.1.Based on the classification and summary of existing literature,the typical architecture of convolutional neural networks and related concepts and classifications of migration learning are sorted out.Several classic convolutional neural network models(Alexnet,VGG,Googlenet,Resnet,and Squeezenet)were selected to analyze their major improvements in detail.2.Six different datasets(Cifar-10,CUB-200-2011,Caltech101,FashionMinist,LZU blade dataset and MEW dataset,and Cifar-10 set four sub-samples with decreasing sample size respectively.)were selected,through experiments to compare the migration ability of different models on different types and different scale data sets.Through comparative experiments,we found that the migration capabilities of the Googlenet and Resnet models have significant advantages over the other two models.Alexnet performs extremely poorly on CUB datasets with small differences between classes and with complex backgrounds.Squeezenet,as a simplified network against Alexnet,has achieved close accuracy with Alexnet on most data sets with high sample quality.In addition,under the same training conditions,the network convergence speed is greatly affected by the specific data set.Although the Resnet network works best on multiple datasets,its convergence rate is the slowest of several models on datasets with small differences between classes.3.Based on the recognition differences of different models of the same category,we use the idea of integrated learning to integrate the migration model.Four integration strategies based on model voting results and category probabilities are proposed and tested in integrated systems of two,three and four models.The experiment found that the strategy based on the probability of TOP1 and the fourth improvement effect is more significant.In most cases,the rate of increase is greater in the test set with a larger number of samples per class(1000 per class,250 test sets).In addition,the integration of migration models that are close to each other is more enhanced.Resnet does not have its own high accuracy when integrated with Alexnet or Squeezenet,but the integration of Googlenet and these two models is still slightly improved.Experiments show that the proposed integration strategy based on migration model is effective and has certain reference value.