Research On Image Classification And Annotation Based On Deep Learning

The rapid development of multimedia technology and the Internet has led to an explosive growth in the amount of image data,which brings great challenges to image retrieval.Although the traditional text-based image retrieval algorithm has high accuracy,it requires manual labeling of images,which is time-consuming and labor-intensive,and cannot meet the retrieval requirements of massive images.Although the automatic image annotation algorithm quickly improves the retrieval efficiency of images,the accuracy of annotation needs to be improved.ML-GCN(Multi-Label Graph Convolutional Networks,Multi-Label Graph Convolutional Networks)has powerful modeling capabilities and non-Euclidean distance calculation capabilities,and can effectively perform operations on multi-label association relationships.In view of this,this article is based on ML-GCN model,transforming the problem of automatic image annotation into the problem of image multi-label classification.Here is main research of this essay:In view of the problem of information loss in image single-label classification and the attention mechanism only using the local correlation between image regions,this essay proposes an improved ML-GCN based on the ML-GCN network model,using the dependencies between multiple labels Image automatic annotation algorithm.The algorithm mainly includes two steps.In the first step,each label node is represented by a word embedding vector.The directed graph constructed by the label correlation matrix is used to model the inter-label dependencies,and the mapping function is used to map the category labels to the corresponding category Classifier;In the second step,the image features extracted by the convolutional neural network are applied to the classifier to obtain the image label.Compared with the Resnet network model,the Resnext network requires less calculation and requires less hyperparameter adjustments.Compared with the ReLU activation function,the mish activation function has the advantages of good training stability,high average accuracy,and high peak accuracy.In view of this,this essay uses the Resnext101 residual network to extract image features,and mish as an activation function proposes an improved ML-GCN image annotation algorithm.The experimental results on the Voc2007 dataset and the coco dataset show that,with ML-GCN,CNN-RNN,Compared with RLSD,DenseNet121,HCP and other methods,the algorithm in this essay effectively improves the average accuracy.In order to reduce the accuracy of low-frequency word labeling caused by imbalanced training samples,this essay introduces a low-frequency feature extraction channel and proposes an improved ML-GCN labeling algorithm based on dual channels.The algorithm combines low-frequency extracted features and global image features to increase the proportion of low-frequency words in the sample,and applying the fused features to an improved ML-GCN label classifier.The experimental results in the Voc2007 data set show that,compared with the improved ML-GCN method,the algorithm in this essay effectively improves the average accuracy and the accuracy of labeling low-frequency words.