The Multi-label Classification Method Of Fusing The Dynamic And Static Label Correlations By GAT

With the continuous progress and development of technology,multi-label image classification has been widely applied in various fields such as communication,unmanned driving,medical image analysis,aerospace,and remote sensing.In order to further improve the accuracy of multi-label classification,this paper aims to address issues such as how to use attention mechanisms to extract image spatial features,how to capture and fuse dynamic and static label correlations to optimize classification results,and how to combine the advantages of the two types of research to propose the multi-label classification method of fusing the dynamic and static label correlations by GAT.Firstly,the multi-label classification method of fusing the dynamic and static label correlations by GAT is proposed,which includes a feature extraction module based on Transformer and a dynamic and static label correlations fusion module based on GAT.Among them,feature extraction module based on Transformer focuses on how to use attention mechanism to extract image spatial features.Transformer is used as the pooling layer of the image reference model to achieve the use of its attention mechanism to extract image spatial features and generate label feature queues;Dynamic and static label correlations fusion module based on GAT focuses on how to capture and fuse dynamic and static label correlations to optimize classification results.The dynamic label correlation capture method based on GAT,the static label correlation capture method based on data statistics,and the mean based fusion method of dynamic and static label correlation are combined to achieve the capture and fusion of dynamic and static label correlations to optimize tag features.Secondly,based on the multi-label classification method of fusing the dynamic and static label correlations by GAT,the classification head EDGE was constructed and applied to the image baseline model TRes Net-L to form a multi-label classification model.Finally,Comparative experiments on three datasets,VOC2007,VOC2012,and COCO2014,were carried out to evaluate the performance of the model.A ablation experiment was also conducted on VOC2007,and the experimental results were analyzed.The experimental results demonstrate the superior performance of the TRes Net-L-EDGE model compared to the comparative model,and also validate the effectiveness of the multi-label classification method of fusing the dynamic and static label correlations by GAT and the multi-label classification header EDGE in improving the accuracy of the benchmark model.