Dynamic Label Bidirectional Fusion Based Cross-modal Multi-label Image Classification

In multi-label image classification, existing methods exhibit significant shortcomings in encoding semantic features from the textual modality and cross-modal feature fusion between image and text modalities. To tackle these challenges, we propose a cross-modal multi-label image classification method based on Dynamic Label Bidirectional Fusion (DLBF). DLBF constructs a dynamic, learnable word embedding space for semantic feature extraction through the projection layer and backpropagation mechanism, eliminating the dependence on dataset co-occurrence statistics and pre-trained word vectors. Regarding cross-modal feature fusion, DLBF leverages cross-attention mechanisms to effectively align features across modalities and enable global bidirectional interaction, thereby achieving a comprehensive fusion of image-text bimodal features. Experimental results on two benchmark datasets demonstrate that the proposed DLBF method fully exploits and integrates image-text multimodal information, accurately modeling discriminative features for classification, thereby improving the accuracy of multi-label image classification tasks.