基于图注意力网络的多标签图像分类模型

针对ML-GCN中标签共现嵌入维度过高影响模型分类性能和ML-GCN中没有充分发掘标签之间不对称关系的问题,提出一种基于图注意力网络的多标签图像分类模型ML-GAT;ML-GAT模型首先对高维标签语义嵌入矩阵进行降维;然后通过降维后的低维标签语义嵌入表示和标签类别共现图得到标签共现嵌入;与此同时ML-GAT将多标签原始...

Multi label Image Classification Model Based on Graph Attention Network

In order to solve the problem that the high co-occurrence dimension of labels in ML-GCN reduces the model classification performance and the asymmetrical relationship between labels is not fully explored in ML-GCN, a multi label image classification model of ML-GAT based on graph attention network is proposed. Firstly, the ML-GAT model reduces the dimensionality of the semantic embedding matrix of high dimensional labels. Then the label co-occurrence embedding is obtained by the low dimensional label semantic embedding representation and the label category co-occurrence graph after dimensionality reduction. At the same time, ML-GAT inputs the original multi label image into the convolutional neural network to extract the general features of the image, and the general features of the multi label image extracted by the convolutional neural network are unified in dimension according to the embedded dimensions of the labels calculated by the graph attention network. Finally, ML-GAT fusion of the image features after co-occurrence and dimensionality reduction of labels is used to obtain the label prediction score of each multi label image. Experimental results on VOC 2007 and MS-COCO 2014 show that ML-GAT achieves good experimental results under the condition of sufficient training samples and sufficient number of label categories. By comparing with other models, the strategy adopted by ML-GAT model can improve the multi label image classification performance of the model to a certain extent.