Prompt-guided consistency learning for multi-label classification with incomplete labels

Shouwen Wang, Qian Wan, Zihan Zhang, Zhigang Zeng

Abstract: Addressing insufficient supervision and improving model generalization are essential for multi-label classification with incomplete annotations, i.e., partial and single positive labels. Recent studies incorporate pseudo-labels to provide additional supervision and enhance model generalization. However, the noise in pseudo-labels generated by the model tends to accumulate, resulting in confirmation bias during training. Self-correction methods, commonly used approaches for mitigating confirmation bias, rely on model predictions but remain susceptible to confirmation bias caused by visual confusion, including both visual ambiguity and similarity. To reduce visual confusion, we propose a prompt-guided consistency learning (PGCL) framework designed for two incomplete labeling settings. Specifically, we introduce an intra-category supervised contrastive loss, which imposes consistency constraints on reliable positive class samples in the feature space of each category, rather than across the feature space of all categories, as in traditional inter-category supervised contrastive loss. Building on this, the distinction between true positive and visual confusion samples for each category is enhanced through label-level contrasting of the same category. Additionally, we develop a class-specific semantic decoupling module that leverages CLIP’s strong vision-language alignment capability, since the proposed contrastive loss requires high-quality label-level representations as contrastive samples. Extensive experimental results on multiple datasets demonstrate that our method can effectively address the problems of two incomplete labeling settings and achieve state-of-the-art performance.