A MULTI-LABEL IMAGE RECOGNITION NETWORK DRIVEN BY LABEL-IMAGE SEMANTIC ALIGNMENT

Abstract

Multi-label image recognition aims to predict a set of semantic labels for an image and has wide applications in other fields. Existing methods have two main problems: the attention regions generated by attention-based methods are insufficiently correlated with label semantics; methods based on label correlation modeling lack dynamic interaction with visual image content, making it difficult to achieve precise alignment between label semantics and image regions. To address these issues, this paper proposes a novel multi-label image recognition algorithm. First, we construct a label graph and leverage a graph convolutional network to learn label semantic priors, modeling the dependencies among labels. Second, we design a semantic decoupling module that adaptively focuses on relevant image regions under the guidance of label semantics to generate label semantic representations. Finally, we introduce a semantic association reasoning module that employs a multi-head self-attention mechanism to dynamically capture semantic correlations among labels, thereby enhancing the discriminative ability of features. Experimental results on the PASCAL VOC 2007 dataset show that our method achieves 95.0% mAP, surpassing existing state-of-the-art methods and improving by 1.6 percentage points over the baseline SSGRL. Ablation studies further validate the effectiveness of each module.