Image Multi-label Learning Based On Correlation Residual Network Tree Model

Image classification is one of the research hotspots in the field of computer vision,and the image multi-label learning method is widely used in image classification tasks.With the advent of the era of big data,image data is growing rapidly.Different fields pay different attention to image data from different perspectives,resulting in only part of semantic labels on images,and the semantic information expressed by images can not be fully reflected,i.e.,image data is missing labels.With the increasing amount of information contained in the image data,the number of corresponding label categories is also increasing,and the set of possible labels predicted for the image samples will grow exponentially,leading to the problem of large prediction output space in the process of multi-label classification.With the development of deep learning,classification models based on convolutional neural networks have powerful representation capabilities and can extract high-level semantic features of images.Compared with traditional classification models,performance and accuracy have been greatly improved.Therefore,this paper aims at the problems of large prediction output space and insufficient training data in the image multi-label classification task with missing labels,and proposes a deep residual network based on convolutional neural network as the backbone network to construct a correlation residual network tree model,Used for classification tasks of multi-label image data.The effectiveness of the model is verified through experiments,and the main research work is as follows:(1)Propose a deep network tree model to classify multi-label image data with missing labels.In the deep network tree model,each deep neural network is regarded as a different branch structure in the model.Then,using different branches of the model to independently train a base classifier for each label category in the image data,and predict each label category attribute of the image sample,thereby avoiding the problem of large prediction output space in the process of multi-label classification.(2)Propose a learning sample selection strategy based on label co-occurrence,which is used to solve the interference caused by insufficient training data and missing label image samples in the training process of the deep network tree model.By calculating the probability that the two label categories are simultaneously associated with an image sample from the original label information in the image data,the co-occurrence between the label categories is described.For each label category,select the image sample corresponding to the label category with strong co-occurrence as the positive example learning sample,and vice versa,as the negative example learning sample,which is the basis of each branch in the deep network tree model.The classifier constructs the corresponding binary training data set for learning.Through experiments on three multi-label image data sets,a deep network tree model based on the Residual Network(Res Net)as the base classifier is obtained.The overall performance is the best.Then the residual network tree model(Res Net-Tree)is proposed.(3)Propose the Correlation Res Net-Tree model(C-RNT),which uses the semantic correlation between label categories to further improve the classification performance of the model.The label correlation matrix is constructed through the semantic correlation between the label categories.Based on the low-rank structural representation of the matrix,the high-order semantic correlation between the label categories is mined,and the low-rank representation matrix is used to enhance the original label matrix,Thus the correlation between the image sample and the label category is modeled,and get the correlation model.The least-square objective loss function is used to jointly learn the Res Net-Tree model with the correlation model,establish a C-RNT model and perform label recovery on image samples.The experimental results show that the C-RNT model proposed in this paper can effectively solve the problems of large prediction output space and insufficient training data in the multi-label process with missing image data labels.It has improved in various performance indicators compared with traditional image multi-label classification algorithms,which have strong theoretical and practical significance for image multi-label learning.