Dual-prompt semantic decoupling and transfer for long-tailed recognition

Real-world data typically exhibit a long-tailed distribution, where a few head classes contain abundant samples while most tail classes suffer from severe data scarcity. This imbalance biases deep models toward head classes, resulting in unreliable representations for tail classes, particularly when class-specific features become entangled with spurious contextual correlations. To address this issue, we propose a Dual-Prompt Semantic Decoupling and Transfer (DSDT) framework for long-tailed recognition. DSDT introduces two complementary visual prompts: a Discriminative Prompt, which emphasizes class-specific cues to enhance inter-class separability, and a Contextual Prompt, which captures shared background context. By explicitly decoupling discriminative and contextual knowledge, the model learns more robust representations for tail classes while mitigating contextual interference. Furthermore, we design a Contextual Prompt Transfer Enhancement mechanism that transfers contextual prototypes from semantically similar head classes to tail classes, enriching contextual diversity and improving representation quality for under-represented categories. Extensive experiments on CIFAR-100-LT, ImageNet-LT and iNaturalist 2018 demonstrate that DSDT consistently outperforms existing methods and significantly improves recognition performance on tail classes.