FAM-Logo: forward compatible multimodal framework for Few-Shot Logo Incremental classification

Logo classification has gained increasing attention for its various applications, including copyright protection, product recommendation and contextual advertising. However, the rapid emergence of new companies and product logo categories presents challenges to existing logo classification models. Often, only a limited number of examples are available for these new categories. A robust classification model should incrementally learn and recognize new logo classes while maintaining its ability to discriminate between existing ones. To address this, we formulate the task of Logo Few-Shot Class-Incremental Learning (Logo-FSCIL). In this work, we propose a new paradigm for Logo-FSCIL. Specifically, we adopt a forward-compatible embedding space reservation strategy and develop a prompt-based multimodal alignment framework to mitigate catastrophic forgetting. Furthermore, taking into account the unique attributes of logos, we devise an enhancement alignment strategy guided by textual information. Experimental results on three logo datasets demonstrate that our method significantly outperforms state-of-the-art FSCIL benchmarks. Extensive experiments on three generic datasets further validate the generalization capability of our approach. Datasets and code can be found at https://github.com/housujuan123/Logo-FSCIL.