The key challenge of counterfeit detection is that as the forgery process continues to upgrade, sample characteristics continue to change, and performance deteriorates when the distribution of training and test data does not match. Our research found that the fundamental reason for performance decline is that the model tends to learn significant but irrelevant background features in the training set. This study addresses the challenge of LV Monogram detection, where existing methods are often affected by background interference and insufficient edge sensitivity. To mitigate these issues, we propose a novel framework that introduces edge-aware feature screening as a complementary modality, enhanced by attention mechanisms and multimodal fusion. Furthermore, a hard sample generation strategy is employed to alleviate model dependence on background cues, thereby enhancing generalization. A custom dataset with varying classification difficulty is constructed for evaluation. Experimental results demonstrate that the proposed method maintains high recognition accuracy across both simple and complex samples, validating its robust feature representation and generalization capability.