Research On Ensemble Methods Of Transfer Learning For Natural Image Recognition

The task of image recognition is mainly based on the feature maps of images.Its purpose is to allow the machine to have ability like human processing image information.On the one hand,it can replace human beings dealing with large number of complicated image data,on the other hand,the computer is easier and more sensitive to deal the pixels of the image information than the human eye.This advantage can be used to solve the difficulty in human processing information.The ability of deep learning makes it well applied in various fields,such as machine vision,Natural Language Processing,Computer Aided Diagnosis and other business areas.With the coming of the era of big data,especially the Image Net data collected by Li Feifei and others,the ILSVRC based on data set makes deep learning develop rapidly in recent years.However,there are still some small sets or inadequate data label prediction classification problems waiting to be solved,such as medical assistant diagnostic tasks.Transfer Learning is a new way to solve the problem of target domain by learning from the original domain.To some extent,it can solve the problem of low accuracy caused by the shortage of data needs for training CNN.The application of Transfer Learning in Deep Learning can enhance the learning ability of neural network,and effectively improve the efficiency of neural network and the ability to recognize image.The Ensemble Learning can steadily improve the recognition efficiency of the convolution neural network.In recent years,the Ensemble Learning has also been applied to various methods of Machine Learning.How to reduce the cost of the Ensemble Learning is also one of the hottest topics of research.Aiming at the problem of natural image recognition,this paper analyzed the detailed network structure of Convolutional Neural Network,and the detailed process of convolution neural network in image feature extraction.A "Ensemble Transfer Learning" method is proposed to solve the problem of prediction and classification of small data sets and insufficient data labels.The main contents and innovation points of this paper are as follows:1.the detailed process of the feature extraction of the Convolution Neural Network in the Image Recognition task and the details of its reverse propagation parameters update are studied.2.studied the ability in Image Recognition of typical Convolutional Neural Network,and their performance are analyzed.Studied and analyzed the how to reduce the cost the Convolutional Neural Network training and the techniques used in improving the image recognition accuracy of training Convolutional Neural Network.3.Combining the method of Ensemble Learning method and transfer convolution neural network parameters,and proposes a new method of Image Recognition,called "Ensemble Transfer learning".Proved this method can improve the recognition accuracy of the powerful ability in reducing the convolutional neural network training cost at the same time through experiment.