Semi-supervised Image Classification Based On Relationship Representation

Semi-supervised learning is a new type of machine learning method between supervised learning and unsupervised learning.Its basic idea is to introduce unlabeled samples in the model training process to solve the problem of model degradation caused by insufficient labeled samples.As deep learning has made major breakthroughs in the field of computer vision,the combination of semi-supervised learning and deep learning has become an inevitable trend,but semi-supervised deep learning methods still face problems such as greater difficulty in constructing regularization items and higher model complexity.In response to the above problems,this paper proposes a semisupervised deep image classification algorithm based on fusion relationship representation from the perspective of relationship learning.The specific content is as follows:First,the existing semi-supervised deep image classification algorithm only revolves around the learning of sample hidden features for modeling,and ignores the implicit relationship between samples,the fusion model of relationship representation and feature extraction which combines the feature extraction of images with the representation of the relationship between images is proposed.The model not only has the ability to extract image features,but also has the ability to extract relationships between images.Finally,numerical experiments show that the proposed model is an image classification method with strong generalization and good classification effect.Secondly,in order to solve the problems of high time complexity and poor adaptive ability when constructing the graph structure algorithm of relational representation,a semi-supervised image classification algorithm based on adaptive relational representation is proposed.The algorithm is still based on the relationship representation and feature extraction fusion model.By replacing the correlation matrix between the images as a dynamic matrix,and adding the optimization of the dynamic matrix to the loss function,it is automatically adjusted along with the optimization process of the model.Realize adaptive dynamic adjustment,so as to avoid the fixed defect of sample relationship in the learning process.Experimental results on MNIST,CIFAR10,CIFAR100,SVHN,STL10 datasets show that the proposed model achieves high classification accuracy,which proves that the deep representation network designed by the mechanism of feature fusion and relational representation fusion is a effective methods to semi-supervised images classification.The paper has 29 pictures,8 tables,and 77 references.