Research On Image Classification Based On Transfer Learning And Deep Convolution Networks

With the development of big data and mobile internet,more and more images appear as floods.Due to the huge number of images,the requirements for image classification models are getting stronger,not only for the complexity of the model,but also for the efficiency of the operation.Therefore,we need to study a method to quickly and efficiently solve the problem of image classification.Transfer learning refers to the transfer of previously learned experience knowledge to new task learning,helping to classify new tasks.It is a way of learning to influence another kind of learning.Transfer learning solves the limitations of training data and test data in traditional machine learning that must obey the same distribution.It can deeply dig into the constant structure and features between the source domain and the target domain,and it can also be highly effective for tasks in the field.The information sharing migration can also be used to migrate and multiplex supervisory information with source domain labels.This paper presents a research method based on transfer learning and deep convolutional network in image classification.This article mainly from a few aspects of research and improvement:(1)For the common image classification,this paper proposes an image classification method based on feature mapping transfer learning,which uses the MK-MMD+joint probability adaptation method to reduce the difference between the target domain and the source domain,and then uses a deep convolutional neural network to extract A highly "enriched" feature vector allows the classifier to more accurately determine the category to which it belongs.MK-MMD evolved from MMD.For the JDA algorithm using MMD+joint probabilistic adaptation,the most important is the nuclear k.Due to the limitation of a single fixed kernel,there exists a Gaussian kernel or linear kernel in practical applications.In the case of selective conditions,MK-MMD is used here,a kernel is constructed with multiple cores with weights,and edge distribution and conditional distribution methods are combined to extract the feature vector with the smallest difference between fields.Further used in the deep learning classification.Experiments show that the use of MK-MMD+joint probabilistic adaptation method significantly improves the feature extraction ability and image classification accuracy.(2)In the case of image data classification,where high-similarity images are prone to misclassification,a transfer learning method using Power Iteration Clustering(PIC)is proposed here to quickly and efficiently isolate similarities.matrix.Divide data sets into ordinary data sets and special data sets.Special data sets are constructed from special data subsets.Then the data set is used to fine-tune the model twice,so that the feature extraction ability and classification performance of the model are further enhanced..Experiments show that using the PIC method has significantly improved the classification effect and speed compared with the spectral clustering method.Finally,selective freezing of the model convolution layer is used,and the classification effect is also improved.