Label correlation and knowledge-guided multilabel image recognition with partial labels

Due to the rich image-to-label correspondence, visual-language pre-trained models have been extensively applied in Multi-Label Image Recognition with Partial Labels (MLIR-PL). However, existing research often overlooks the correlations between labels, which limits the model's ability to handle label co-occurrence relationships effectively. Additionally, the prompts learned from unsupervised data frequently fail to accurately capture the attributes of categories. To tackle these issues, we proposed a new MLIR-PL method, LCKG. Specifically, LCKG employs graph neural networks (GNNs) to learn label correlations and introduces similarity loss to mitigate modality discrepancies. By combining knowledge-guided methods, this study optimizes the information transfer between handcrafted prompts and learnable prompts, enabling the learnable prompts to more accurately describe category attributes. Experimental results on the COCO 2014 and VOC 2007 datasets demonstrate the superior performance of our proposed method.