| Multi-label image classification is a research hotspot in the problem of multi-label data classification,and multi-label images have been widely studied and applied in communications,unmanned driving,medical image analysis,aerospace,remote sensing and other fields,and play an important role in people’s lives and national economy.How to efficiently use the label co-occurrence relationship in multi-label images to improve the classification performance of multi-label image classification model is a meaningful research work.In view of the existing problems and important application value of the current multi-label image classification method,this paper combines the attention mechanism of multi-label graph and the graph model to propose two multi-label image classification models:(1)Aiming at the problem that the current multi-label image classification method only learns the visual representation features of images and ignores the semantic feature information of image labels,a multi-label image classification model(ML-M-GAT)based on multi-head graph attention network and graph model is proposed.The model includes image feature extraction module based on Res Net101 model,label vector transformation module based on word embedding and graph structure,label attention weight learning module based on multi-head graph attention mechanism,and classifier module based on fusion features.In experiments on the public datasets VOC-2007 and COCO-2014,the average mean accuracy(m AP)of the ML-M-GAT model was 94% and 82.2%,respectively,which were better than the CNN-RNN,Res Net101,MLIR,MIC-FLC models,and improved by 4.2% and 3.9% compared with the Res Net101 model,respectively.Experimental results show that the proposed ML-M-GAT model can improve the performance of multi-label image classification by using image label information.(2)In order to further improve the accuracy of multi-label image classification,this paper introduces the channel attention mechanism and spatial attention mechanism into the image visual feature extractor,learns the features of important parts in the image through CBAM,and proposes a multi-label image classification model(V-ML-M-GAT)based on the text-visual multi-head graph attention and graph model.Based on the deep neural network extracting image features,the model introduces the multi-head text-visual attention mechanism and graph model to improve the performance of the multi-label image classification model.The average mean accuracy(m AP)of the V-ML-M-GAT model on the VOC-2007 and COCO-2014 datasets was 95.7% and 84.3%,respectively,which were 1.7%and 2.1% higher than the ML-M-GAT model.Experimental results show that the proposed V-ML-M-GAT model can better learn the visual representation features of multi-label images,and can also learn the relevant information between image labels and the correspondence between label semantics and image features,and has good multi-label image classification performance. |
