IKA²: Internal Knowledge Adaptive Activation for Robust Recognition in Complex ScenariosIKA²: Internal knowledge adaptive activation for robust recognition in complex scenarios

Confusing image classification aims to address both inter-instance similarity interference and instance-scene similarity interference. However, existing methods typically treat these two challenges in isolation, leading to fragmented research progress. To this end, we propose a novel plug-and-play module, termed internal knowledge adaptive activation (IKA²), which is designed to activate and exploit the latent knowledge embedded in large-scale pre-trained vision-language models, thereby improving performance on tasks prone to category confusion. It consists of two key components: Knowledge completion (KC), which learns task-relevant prompt vectors to dynamically activate dormant knowledge associated with easily confusable categories; and irrelevant suppression (IS), which implements prompting at the image level, suppresses confusing background and noise, and dynamically guides the model to focus on the core target object. Additionally, we introduce a feature activation guidance loss to maximize inter-class separability in the feature space, thus enhancing the model′s discriminative capability. Comprehensive experiments conducted on three challenging datasets across two representative scenarios demonstrate that IKA² consistently outperforms state-of-the-art methods, achieving significant improvements in classification accuracy and generalization capability. These results highlight the effectiveness and transferability of IKA², making it a promising solution for robust image classification in complex real-world environments.