Supervised Contrastive Learning and Dynamic Attention Fusion for Integrating Deep and Handcrafted Features in Anti-counterfeiting

This study proposes an intelligent framework based on multimodal texture features analysis to address the persistent challenges of easy replication and low detection accuracy in traditional product anti-counterfeiting systems. The proposed approach integrates an enhanced supervised contrastive learning architecture with differentiable handcrafted feature extraction to construct a highly discriminative model that enhances both deep and handcrafted representations. A novel hybrid contrastive learning mechanism—combining InfoNCE loss with triplet sampling and explicitly masking same-class interfering samples—is introduced to optimize the feature-space distribution. Furthermore, Local Binary Patterns (LBP) are reformulated as differentiable convolutional operations to enable multiscale spatial pyramid fusion. A dynamic attention module is introduced to effectively fuse deep features extracted via RepVGG with multiscale LBP handcrafted features. Experimental evaluations on the ALOT dataset demonstrate that the proposed framework achieves a classification accuracy of 99.71% in closed-set scenarios and 97.43% in open-set scenarios, effectively fulfilling stringent anti-counterfeiting requirements across diverse conditions and offering a robust, efficient solution for product authentication.