Multi-Label Image Classification Based On Multi Attention

Multi-label image classification is an important topic in the field of computer vision and pattern recognition.Because in the real world,objects such as images,videos,music,documents and so on are usually described by multi-label attributes,multi-label image classification is a more practical and complex task compared with single label image classification.Multi label image classification usually has more complex output space.Besides,in image features,the corresponding features of each label often interfere with each other in the recognition process.In order to solve the problem of multi label image classification,a multi-label image classification method based on multi attention mechanism(MAML)is proposed in this paper.Multiple attention means that two different attention models are used in this model which are named as class activation mapping and cross modality multi head spatial attention.MAML uses the graph convolution neural network for label enhancement learning,then the representation of labels and the image features are used to calculate the cross-modality spatial attention heatmap.Through the label embedding representation obtained by the graph convolution network,the relationship information between labels is implicitly included in MAML.Because the label embeddings are used as the weight matrix of class activation mapping,so that the information of label association can be transferred in the network.In this way,the label dependency information learned from convolution can be fully utilized.Finally,MAML is evaluated on three standard multi label image data sets MS-COCO,Pascal VOC 2007 and NUS-WIDE.The experimental results show that the effectiveness of MAML is better than or similar to the existing multi label image classification methods.In addition,the ablation experiments were carried out on some parameters in the experiment,and the influence of parameter changes on the MAML method are also analyzed.